A Spark DataSouce implementation for reading files into Arrow compatible columnar vectors.
The development of this library is still in progress. As a result some of the functionality may not be constantly stable for being used in production environments that have not been fully considered due to the limited testing capabilities so far.
There are some requirements before you build the project. Please make sure you have already installed the software in your system.
- gcc 9.3 or higher version
- java8 OpenJDK -> yum install java-1.8.0-openjdk
- cmake 3.2 or higher version
- maven 3.1.1 or higher version
- Hadoop 2.7.5 or higher version
- Spark 3.0.0 or higher version
// installing gcc 9.3 or higher version Please notes for better performance support, gcc 9.3 is a minimal requirement with Intel Microarchitecture such as SKYLAKE, CASCADELAKE, ICELAKE. https://gcc.gnu.org/install/index.html Follow the above website to download gcc. You may have to launch ./contrib/download_prerequisites command to install all the prerequisites for gcc. If you are facing downloading issue in download_prerequisites command, you can try to change ftp to http.
//Follow the steps to configure gcc https://gcc.gnu.org/install/configure.html If you are facing a multilib issue, you can try to add --disable-multilib parameter in ../configure
//Follow the steps to build gcc https://gcc.gnu.org/install/build.html
//Follow the steps to install gcc https://gcc.gnu.org/install/finalinstall.html
//Set up Environment for new gcc export PATH=$YOUR_GCC_INSTALLATION_DIR/bin:$PATH export LD_LIBRARY_PATH=$YOUR_GCC_INSTALLATION_DIR/lib64:$LD_LIBRARY_PATH Please remember to add and source the setup in your environment files such as /etc/profile or /etc/bashrc
//Verify if gcc has been installation Use gcc -v command to verify if your gcc version is correct.(Must larger than 9.3)
If you are facing some trouble when installing cmake, please follow below steps to install cmake.
// installing cmake 3.2
sudo yum install cmake3
// If you have an existing cmake, you can use below command to set it as an option within alternatives command
sudo alternatives --install /usr/local/bin/cmake cmake /usr/bin/cmake 10 --slave /usr/local/bin/ctest ctest /usr/bin/ctest --slave /usr/local/bin/cpack cpack /usr/bin/cpack --slave /usr/local/bin/ccmake ccmake /usr/bin/ccmake --family cmake
// Set cmake3 as an option within alternatives command
sudo alternatives --install /usr/local/bin/cmake cmake /usr/bin/cmake3 20 --slave /usr/local/bin/ctest ctest /usr/bin/ctest3 --slave /usr/local/bin/cpack cpack /usr/bin/cpack3 --slave /usr/local/bin/ccmake ccmake /usr/bin/ccmake3 --family cmake
// Use alternatives to choose cmake version
sudo alternatives --config cmake
If you are facing some trouble when installing maven, please follow below steps to install maven
// installing maven 3.6.3
Go to https://maven.apache.org/download.cgi and download the specific version of maven
// Below command use maven 3.6.3 as an example
wget https://ftp.wayne.edu/apache/maven/maven-3/3.6.3/binaries/apache-maven-3.6.3-bin.tar.gz
tar xzf apache-maven-3.6.3-bin.tar.gz
mkdir /usr/local/maven
mv apache-maven-3.6.3/ /usr/local/maven/
// Set maven 3.6.3 as an option within alternatives command
sudo alternatives --install /usr/bin/mvn mvn /usr/local/maven/apache-maven-3.6.3/bin/mvn 1
// Use alternatives to choose mvn version
sudo alternatives --config mvn
Please make sure you have set up Hadoop directory properly with Hadoop Native Libraries
By default, Apache Arrow would scan $HADOOP_HOME and find the native Hadoop library libhdfs.so(under $HADOOP_HOME/lib/native directory) to be used for Hadoop client.
You can also use ARROW_LIBHDFS_DIR to configure the location of libhdfs.so if it is installed in other directory than $HADOOP_HOME/lib/native
If your SPARK and HADOOP are separated in different nodes, please find libhdfs.so in your Hadoop cluster and copy it to SPARK cluster, then use one of the above methods to set it properly.
For more information, please check Arrow HDFS interface documentation Hadoop Native Library, please read the official Hadoop website documentation
For better performance ArrowDataSource reads HDFS files using the third-party library libhdfs3. The library must be pre-installed on machines Spark Executor nodes are running on.
To install the library, use of Conda is recommended.
// installing libhdfs3
conda install -c conda-forge libhdfs3
// check the installed library file
ls -l ~/miniconda/envs/$(YOUR_ENV_NAME)/lib/libhdfs3.so/lib/libhdfs3.so
To set up libhdfs3, there are two different ways:
Option1: Overwrite the soft link for libhdfs.so
To install libhdfs3.so, you have to create a soft link for libhdfs.so in your Hadoop directory($HADOOP_HOME/lib/native by default).
ln -f -s libhdfs3.so libhdfs.so
Option2: Add env variable to the system
export ARROW_LIBHDFS3_DIR="PATH_TO_LIBHDFS3_DIR/"
Add following Spark configuration options before running the DataSource to make the library to be recognized:
spark.executorEnv.ARROW_LIBHDFS3_DIR = "PATH_TO_LIBHDFS3_DIR/"spark.executorEnv.LD_LIBRARY_PATH = "PATH_TO_LIBHDFS3_DEPENDENCIES_DIR/"
Please notes: If you choose to use libhdfs3.so, there are some other dependency libraries you have to installed such as libprotobuf or libcrypto.
You have to use a cusotmized Arrow to support for our datasets Java API.
// build arrow-cpp
git clone --branch native-sql-engine-clean https://github.com/Intel-bigdata/arrow.git
cd arrow/cpp
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DARROW_DEPENDENCY_SOURCE=BUNDLED -DARROW_PARQUET=ON -DARROW_HDFS=ON -DARROW_BOOST_USE_SHARED=ON -DARROW_JNI=ON -DARROW_WITH_SNAPPY=ON -DARROW_WITH_PROTOBUF=ON -DARROW_DATASET=ON ..
make
// build and install arrow jvm library
cd ../../java
mvn clean install -P arrow-jni -am -Darrow.cpp.build.dir=$PATH_TO_ARROW_SOURCE_CODE/arrow/cpp/build/release
// Download OAP Source Code, replace $BRANCH_NAME to the target branch you want to download.
git clone --branch $BRANCH_NAME https://github.com/Intel-bigdata/OAP.git
// Go to the directory
cd $PATH_TO_OAP_DIR/oap-data-source/arrow
// build
mvn clean package
// check built jar library
readlink -f standard/target/spark-arrow-datasource-standard-0.9.0-jar-with-dependencies.jar
Currently ArrowDataSource works on the Spark 3.0.0 version.
wget http://archive.apache.org/dist/spark/spark-3.0.0/spark-3.0.0-bin-hadoop2.7.tgz
tar -xf ./spark-3.0.0-bin-hadoop2.7.tgz
export SPARK_HOME=`pwd`/spark-3.0.0-bin-hadoop2.7
If you are new to Apache Spark, please go though Spark's official deploying guide before getting started with ArrowDataSource.
To enable ArrowDataSource, the previous built jar spark-arrow-datasource-standard-0.9.0-jar-with-dependencies.jar should be added to Spark configuration. Typically the options are:
spark.driver.extraClassPath: Set to load jar file to driver.spark.executor.extraClassPath: Set to load jar file to executor.jars: Set to copy jar file to the executors when using yarn cluster mode.spark.executorEnv.ARROW_LIBHDFS3_DIR: Optional if you are using a custom libhdfs3.so.spark.executorEnv.LD_LIBRARY_PATH: Optional if you are using a custom libhdfs3.so.
Example to run Spark Shell with ArrowDataSource jar file
${SPARK_HOME}/bin/spark-shell \
--verbose \
--master yarn \
--driver-memory 10G \
--conf spark.driver.extraClassPath=/path-to-jar-dir/spark-arrow-datasource-standard-0.9.0-jar-with-dependencies.jar \
--conf spark.executor.extraClassPath=/path-to-jar-dir/spark-arrow-datasource-standard-0.9.0-jar-with-dependencies.jar \
--conf spark.driver.cores=1 \
--conf spark.executor.instances=12 \
--conf spark.executor.cores=6 \
--conf spark.executor.memory=20G \
--conf spark.memory.offHeap.size=80G \
--conf spark.task.cpus=1 \
--conf spark.locality.wait=0s \
--conf spark.sql.shuffle.partitions=72 \
--conf spark.executorEnv.ARROW_LIBHDFS3_DIR=/path-to-libhdfs3-dir/ \
--conf spark.executorEnv.LD_LIBRARY_PATH=/path-to-libhdfs3-dir/ \
--jars /path-to-jar-dir/spark-arrow-datasource-0.9.0-jar-with-dependencies.jar
For more information about these options, please read the official Spark documentation.
val path = "${PATH_TO_YOUR_PARQUET_FILE}"
val df = spark.read
.option(ArrowOptions.KEY_ORIGINAL_FORMAT, "parquet")
.option(ArrowOptions.KEY_FILESYSTEM, "hdfs")
.format("arrow")
.load(path)
df.createOrReplaceTempView("my_temp_view")
spark.sql("SELECT * FROM my_temp_view LIMIT 10").show(10)To validate if ArrowDataSource works, you can go to the DAG to check if ArrowScan has been used from the above example query.
We provide a customized replacement of Spark's built-in ParquetFileFormat. By so users don't have to change existing Parquet-based SQL/code and will be able to read Arrow data from Parquet directly. More importantly, sometimes the feature could be extremely helpful to make ArrowDataSource work correctly with some 3rd-party storage tools (e.g. Delta Lake) that are built on top of ParquetDataSource.
To replace built-in ParquetDataSource, the only thing has to be done is to place compiled jar spark-arrow-datasource-parquet-0.9.0.jar into
Spark's library folder.
If you'd like to verify that ParquetDataSource is successfully overwritten by the jar, run following code before executing SQL job:
ServiceLoaderUtil.ensureParquetFileFormatOverwritten();
Note the whole feature is currently experimental and only DataSource v1 is supported. V2 support is being planned.