Clean up CI ASAN & .bazelrc (#8828)

.bazelrc

@@ -31,6 +31,7 @@ build --host_copt="-Wno-microsoft-unqualified-friend"
 # This workaround is needed due to https://github.com/bazelbuild/bazel/issues/4341
 build --per_file_copt="-\\.(asm|S)$,external/com_github_grpc_grpc/.*@-DGRPC_BAZEL_BUILD"
 build --http_timeout_scaling=5.0
+build --verbose_failures
 build:iwyu --experimental_action_listener=//:iwyu_cpp
 
 # Print relative paths when possible
@@ -62,7 +63,6 @@ build:tsan --copt -fno-omit-frame-pointer
 build:tsan --linkopt -fsanitize=thread
 
 # Memory sanitizer configuration
-# If you change these make sure you also change ci/asan_tests/bazelrc
 build:asan --strip=never
 build:asan --copt -g
 build:asan --copt -fsanitize=address
@@ -71,17 +71,27 @@ build:asan --copt -DADDRESS_SANITIZER
 build:asan --copt -fno-omit-frame-pointer
 build:asan --linkopt -fsanitize=address
 build:asan --linkopt -fno-sanitize=vptr
-test:asan --action_env=ASAN_OPTIONS=detect_leaks=0
+test:asan --jobs=1
+test:asan --test_env=ASAN_OPTIONS="detect_leaks=0"
 # This LD_PRELOAD is set for Travis. You will need to change it for local debugging.
-test:asan --action_env=LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libasan.so.2
+test:asan --test_env=LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libasan.so.2 /usr/lib/gcc/x86_64-linux-gnu/7/libasan.so"
 # For example, for Ubuntu 18.04 libasan can be found here:
-# test:asan --action_env=LD_PRELOAD=/usr/lib/gcc/x86_64-linux-gnu/7/libasan.so
+# test:asan --test_env=LD_PRELOAD="/usr/lib/gcc/x86_64-linux-gnu/7/libasan.so"
 
+# CI configuration:
+build:ci --color=yes
+build:ci --curses=no
+build:ci --disk_cache=~/ray-bazel-cache
+build:ci --remote_cache="https://storage.googleapis.com/ray-bazel-cache"
+build:ci --progress_report_interval=100
+build:ci --show_progress_rate_limit=15
+build:ci --show_task_finish
+build:ci --ui_actions_shown=1024
+build:ci-travis --show_timestamps  # Travis doesn't have an option to show timestamps, but GitHub Actions does
+# GitHub Actions has low disk space, so prefer hardlinks there.
+build:ci-github --experimental_repository_cache_hardlinks
 test:ci --flaky_test_attempts=3
 test:ci --nocache_test_results
-test:ci --progress_report_interval=100
-test:ci --show_progress_rate_limit=100
-test:ci --show_timestamps
 test:ci --spawn_strategy=local
 test:ci --test_output=errors
 test:ci --test_verbose_timeout_warnings

.travis.yml

@@ -33,7 +33,7 @@ matrix:
 
     - os: linux
       env:
-        - PYTHON=3.6 ENABLE_ASAN="--config=asan -j 2"
+        - PYTHON=3.6 BAZEL_CONFIG="asan"
         - PYTHONWARNINGS=ignore
         - RAY_DEFAULT_BUILD=1
       install:
@@ -42,7 +42,7 @@ matrix:
         - . ./ci/travis/ci.sh build
       script:
         # Run all C++ unit tests with ASAN enabled. ASAN adds too much overhead to run Python tests.
-        - bazel test $ENABLE_ASAN --config=ci --build_tests_only -- //:all
+        - bazel test --config=ci --build_tests_only -- //:all
 
     - os: osx
       osx_image: xcode7
@@ -176,7 +176,7 @@ matrix:
       before_script:
         - . ./ci/travis/ci.sh build
       script:
-        - ./ci/keep_alive bazel test --config=ci --test_output=errors --build_tests_only --test_tag_filters=learning_tests_tf rllib/...
+        - ./ci/keep_alive bazel test --config=ci --build_tests_only --test_tag_filters=learning_tests_tf rllib/...
 
     # RLlib: Learning tests with tf=1.x (from rllib/tuned_examples/*.yaml).
     # Requested by Edi (MS): Test all learning capabilities with tf1.x
@@ -193,7 +193,7 @@ matrix:
       before_script:
         - . ./ci/travis/ci.sh build
       script:
-        - ./ci/keep_alive bazel test --config=ci --test_output=errors --build_tests_only --test_tag_filters=learning_tests_tf rllib/...
+        - ./ci/keep_alive bazel test --config=ci --build_tests_only --test_tag_filters=learning_tests_tf rllib/...
 
     # RLlib: Learning tests with torch (from rllib/tuned_examples/*.yaml).
     - os: linux
@@ -209,7 +209,7 @@ matrix:
       before_script:
         - . ./ci/travis/ci.sh build
       script:
-        - ./ci/keep_alive bazel test --config=ci --test_output=errors --build_tests_only --test_tag_filters=learning_tests_torch rllib/...
+        - ./ci/keep_alive bazel test --config=ci --build_tests_only --test_tag_filters=learning_tests_torch rllib/...
 
     # RLlib: Quick Agent train.py runs (compilation & running, no(!) learning).
     # Agent single tests (compilation, loss-funcs, etc..).

build.sh

@@ -128,7 +128,7 @@ pushd "$BUILD_DIR"
 
 
 if [ "$RAY_BUILD_JAVA" == "YES" ]; then
-  "$BAZEL_EXECUTABLE" build ${ENABLE_ASAN-} //java:ray_java_pkg --verbose_failures
+  "$BAZEL_EXECUTABLE" build //java:ray_java_pkg
 fi
 
 if [ "$RAY_BUILD_PYTHON" == "YES" ]; then
@@ -160,7 +160,7 @@ if [ "$RAY_BUILD_PYTHON" == "YES" ]; then
 
   export PYTHON3_BIN_PATH="$PYTHON_EXECUTABLE"
 
-  "$BAZEL_EXECUTABLE" build ${ENABLE_ASAN-} //:ray_pkg --verbose_failures
+  "$BAZEL_EXECUTABLE" build //:ray_pkg
 fi
 
 popd

ci/asan_tests/Makefile

@@ -1,28 +0,0 @@
-SHELL=/bin/bash
-
-echo:
-	@echo "Available commands: setup, run, recompile"
-
-setup:
-	echo "Setting up the environment"
-	pip install -r ray-project/requirements.txt
-	pip install -U pytest
-
-	echo "Installing cython example"
-	cd "${HOME}"/ray/doc/examples/cython; python setup.py install --user
-
-	echo "Settting up the shell"
-	cp bazelrc "${HOME}"/.bazelrc # Setup cache
-	echo "LD_PRELOAD=/usr/lib/gcc/x86_64-linux-gnu/7/libasan.so" >> "${HOME}"/.bashrc
-	echo "ASAN_OPTIONS=detect_leaks=0" >> "${HOME}"/.bashrc
-
-	echo "Compiling ray"
-	cd "${HOME}"/ray; git fetch; git pull origin master
-	cd "${HOME}"/ray/python; pip install -e . --verbose || true
-
-run:
-	./run_asan_tests.sh
-
-recompile:
-	cd "${HOME}"/ray; git fetch; git checkout "${git_sha}"
-	cd "${HOME}"/ray/python; pip install -e . --verbose || true

ci/asan_tests/bazelrc

@@ -1,30 +0,0 @@
-# This bazelrc specifies configuration for the release asan test.
-# It exists because we want everything to be "default" configuration
-# instead of "--config=ci" or "--config=asan".
-
-# Cache
-build --remote_cache=https://storage.googleapis.com/ray-bazel-cache
-build --disk_cache=/tmp/bazel-cache
-build --repository_cache=/tmp/bazel-cache
-build --remote_upload_local_results=false
-
-# Asan
-build --strip=never
-build --copt -g
-build --copt -fsanitize=address
-build --copt -fno-sanitize=vptr
-build --copt -DADDRESS_SANITIZER
-build --copt -fno-omit-frame-pointer
-build --linkopt -fsanitize=address
-build --linkopt -fno-sanitize=vptr
-test --action_env=ASAN_OPTIONS=detect_leaks=0
-test --action_env=LD_PRELOAD=/usr/lib/gcc/x86_64-linux-gnu/7/libasan.so
-
-# Travis test config
-test --flaky_test_attempts=3
-test --nocache_test_results
-test --show_timestamps
-test --spawn_strategy=local
-test --test_output=errors
-test --test_verbose_timeout_warnings
-test --jobs 1

ci/asan_tests/ray-project/project.yaml

@@ -16,12 +16,12 @@ cluster:
 
 commands:
   - name: Setup
-    command: make setup
+    command: ./run_asan_tests.sh setup
   - name: Run
-    command: make run
+    command: ./run_asan_tests.sh run
   - name: Recompile
     command: |
-      git_sha={{git_sha}} make recompile
+      git_sha={{git_sha}} ./run_asan_tests.sh recompile
     params:
     - name: git_sha # Ray version string.
       default: "master"

ci/asan_tests/run_asan_tests.sh

@@ -1,15 +1,58 @@
-set -e
-set -x
+#!/usr/bin/env bash
 
-export LD_PRELOAD=/usr/lib/gcc/x86_64-linux-gnu/7/libasan.so
-export ASAN_OPTIONS=detect_leaks=0
+set -euxo pipefail
 
-cd $HOME/ray
-# async plasma test
-python -m pytest -v --durations=5 --timeout=300 python/ray/experimental/test/async_test.py
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE:-$0}")"; pwd)"
 
-# Ray tests
-bazel test --test_tag_filters=-jenkins_only python/ray/serve/...
-bazel test --test_tag_filters=-jenkins_only python/ray/dashboard/...
-bazel test --test_tag_filters=-jenkins_only python/ray/tests/...
-bazel test --test_tag_filters=-jenkins_only python/ray/tune/...
+asan_install() {
+  (cd "${ROOT_DIR}"/../python && pip install -e . --verbose)
+}
+
+asan_setup() {
+  echo "Setting up the environment"
+  pip install -r ray-project/requirements.txt
+  pip install -U pytest
+
+  echo "Installing cython example"
+  (cd "${ROOT_DIR}"/../doc/examples/cython && python setup.py install --user)
+
+  echo "Settting up the shell"
+  echo "build --config=asan" >> ~/.bazelrc  # Setup cache
+  echo "LD_PRELOAD=/usr/lib/gcc/x86_64-linux-gnu/7/libasan.so" >> ~/.bashrc
+  echo "ASAN_OPTIONS=detect_leaks=0" >> ~/.bashrc
+
+  echo "Compiling ray"
+  git fetch
+  git pull origin master
+  asan_install || true
+}
+
+asan_run() {
+  (
+    export LD_PRELOAD="/usr/lib/gcc/x86_64-linux-gnu/7/libasan.so"
+    export ASAN_OPTIONS="detect_leaks=0"
+
+    cd "${ROOT_DIR}"/../..
+
+    # async plasma test
+    python -m pytest -v --durations=5 --timeout=300 python/ray/experimental/test/async_test.py
+
+    # Ray tests
+    bazel test --test_tag_filters=-jenkins_only python/ray/serve/...
+    bazel test --test_tag_filters=-jenkins_only python/ray/dashboard/...
+    bazel test --test_tag_filters=-jenkins_only python/ray/tests/...
+    bazel test --test_tag_filters=-jenkins_only python/ray/tune/...
+  )
+}
+
+asan_recompile() {
+  git fetch
+  git checkout "${git_sha}"
+  asan_install || true
+}
+
+if [ 0 -lt "$#" ]; then
+  asan_"$@"
+else
+  echo "Available commands: setup, run, recompile"
+fi

ci/travis/ci.sh

@@ -66,7 +66,7 @@ reload_env() {
     export TRAVIS_PULL_REQUEST
   fi
 
-  if [ -z "${TRAVIS_BRANCH-}" ] && [ -n "${GITHUB_WORKFLOW-}" ]; then
+  if [ "${GITHUB_ACTIONS-}" = true ] && [ -z "${TRAVIS_BRANCH-}" ]; then
     # Define TRAVIS_BRANCH to make Travis scripts run on GitHub Actions.
     TRAVIS_BRANCH="${GITHUB_BASE_REF:-${GITHUB_REF}}"  # For pull requests, the base branch name
     TRAVIS_BRANCH="${TRAVIS_BRANCH#refs/heads/}"  # Remove refs/... prefix
@@ -403,7 +403,7 @@ _check_job_triggers() {
   fi
 
   if ! (set +x && should_run_job ${job_names//,/ }); then
-    if [ -n "${GITHUB_WORKFLOW-}" ]; then
+    if [ "${GITHUB_ACTIONS-}" = true ]; then
       # If this job is to be skipped, emit 'exit' into .bashrc to quickly exit all following steps.
       # This isn't needed on Travis (since everything runs in one shell), but is on GitHub Actions.
       cat <<EOF1 >> ~/.bashrc
@@ -448,7 +448,8 @@ init() {
 
 build() {
   if ! need_wheels; then
-    bazel build ${ENABLE_ASAN-} -k "//:*"   # Do a full build first to ensure everything passes
+    # NOTE: Do not add build flags here. Use .bazelrc and --config instead.
+    bazel build -k "//:*"  # Full build first, since pip install will build only a subset of targets
     install_ray
     if [ "${LINT-}" = 1 ]; then
       # Try generating Sphinx documentation. To do this, we need to install Ray first.
@@ -470,7 +471,7 @@ build() {
 }
 
 _main() {
-  if [ -n "${GITHUB_WORKFLOW-}" ]; then
+  if [ "${GITHUB_ACTIONS-}" = true ]; then
     exec 2>&1  # Merge stdout and stderr to prevent out-of-order buffering issues
     reload_env
   fi

ci/travis/install-bazel.sh

@@ -53,7 +53,7 @@ else
   target="./install.sh"
   curl -f -s -L -R -o "${target}" "https://github.com/bazelbuild/bazel/releases/download/${version}/bazel-${version}-installer-${platform}-${achitecture}.sh"
   chmod +x "${target}"
-  if [ "${TRAVIS-}" = true ] || [ -n "${GITHUB_WORKFLOW-}" ]; then
+  if [ "${CI-}" = true ]; then
     sudo "${target}" > /dev/null  # system-wide install for CI
     command -V bazel 1>&2
   else
@@ -62,36 +62,26 @@ else
   rm -f "${target}"
 fi
 
+for bazel_cfg in ${BAZEL_CONFIG-}; do
+  echo "build --config=${bazel_cfg}" >> ~/.bazelrc
+done
 if [ "${TRAVIS-}" = true ]; then
-  # Use bazel disk cache if this script is running in Travis.
-  cat <<EOF >> "${HOME}/.bazelrc"
-build --show_timestamps  # Travis doesn't have an option to show timestamps, but GitHub Actions does
-# If we are in Travis, most of the compilation result will be cached.
-# This means we are I/O bounded. By default, Bazel set the number of concurrent
-# jobs to the the number cores on the machine, which are not efficient for
-# network bounded cache downloading workload. Therefore we increase the number
-# of jobs to 50
-build --jobs=50
-EOF
+  echo "build --config=ci-travis" >> ~/.bazelrc
+
+  # If we are in Travis, most of the compilation result will be cached.
+  # This means we are I/O bounded. By default, Bazel set the number of concurrent
+  # jobs to the the number cores on the machine, which are not efficient for
+  # network bounded cache downloading workload. Therefore we increase the number
+  # of jobs to 50
+  # NOTE: Normally --jobs should be under 'build:ci-travis' in .bazelrc, but we put
+  # it under 'build' here avoid conflicts with other --config options.
+  echo "build --jobs=50" >> ~/.bazelrc
 fi
-if [ -n "${GITHUB_WORKFLOW-}" ]; then
-  cat <<EOF >> "${HOME}/.bazelrc"
---output_base=".bazel-out"  # On GitHub Actions, staying on the same volume seems to be faster
-EOF
+if [ "${GITHUB_ACTIONS-}" = true ]; then
+  echo "build --config=ci-github" >> ~/.bazelrc
 fi
-if [ "${TRAVIS-}" = true ] || [ -n "${GITHUB_WORKFLOW-}" ]; then
-  cat <<EOF >> "${HOME}/.bazelrc"
-# CI output doesn't scroll, so don't use curses
-build --color=yes
-build --curses=no
-build --disk_cache="$(test "${OSTYPE}" = msys || echo ~/ray-bazel-cache)"
-# Use ray google cloud cache
-build --remote_cache="https://storage.googleapis.com/ray-bazel-cache"
-build --show_progress_rate_limit=15
-build --show_task_finish
-build --ui_actions_shown=1024
-build --verbose_failures
-EOF
+if [ "${CI-}" = true ]; then
+  echo "build --config=ci" >> ~/.bazelrc
   # If we are in master build, we can write to the cache as well.
   upload=0
   if [ "${TRAVIS_PULL_REQUEST-false}" = false ]; then
@@ -113,16 +103,16 @@ EOF
     fi
   fi
   if [ 0 -ne "${upload}" ]; then
-    translated_path="${HOME}/bazel_cache_credential.json"
+    translated_path=~/bazel_cache_credential.json
     if [ "${OSTYPE}" = msys ]; then  # On Windows, we need path translation
       translated_path="$(cygpath -m -- "${translated_path}")"
     fi
-    cat <<EOF >> "${HOME}/.bazelrc"
+    cat <<EOF >> ~/.bazelrc
 build --google_credentials="${translated_path}"
 EOF
   else
     echo "Using remote build cache in read-only mode." 1>&2
-    cat <<EOF >> "${HOME}/.bazelrc"
+    cat <<EOF >> ~/.bazelrc
 build --remote_upload_local_results=false
 EOF
   fi

ci/travis/install-dependencies.sh

@@ -162,7 +162,7 @@ install_pip() {
     "${python}" -m pip install --upgrade --quiet pip
 
     # If we're in a CI environment, do some configuration
-    if [ "${TRAVIS-}" = true ] || [ -n "${GITHUB_WORKFLOW-}" ]; then
+    if [ "${CI-}" = true ]; then
       "${python}" -W ignore -m pip config -q --user set global.disable-pip-version-check True
       "${python}" -W ignore -m pip config -q --user set global.no-color True
       "${python}" -W ignore -m pip config -q --user set global.progress_bar off