Spark on Docker on Yarn on Azure HDInsight
Feb 18, 2023
devtip , docker , yarn , spark , azure , hdinsight
This post builds on Docker on Yarn on Azure HDInsight and covers tricks to run Spark workloads on Docker on Yarn on Azure HDInsight
We will first run pi
example on Spark followed by a sample numpy
script inside docker containers.
Pre-requisite
- Create an HDInsight cluster (Spark 3.1) and one worker node.
PS: Keeping it to one worker node gives more control on where the containers get created and that means it's easier to debug should things go wrong.
All the steps in this post, can be completed from the workernode itself. First ssh to one of the headnodes and from there ssh to the workernode to follow the steps below
Prepare the cluster for running Docker workload
Install Docker on the worker node.
Run following script either using custom script action or by sshing to the node directly as $ sudo bash -x install_docker.sh
#!/bin/bash # install_docker.sh set -e sudo apt-get -y remove docker docker-engine docker.io containerd runc sudo apt-get -y update sudo apt-get -y install \ ca-certificates \ curl \ gnupg \ lsb-release sudo rm -rf /etc/docker /var/lib/docker /run/docker.sock /etc/systemd/system/docker.service.d curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo gpg --dearmor -o /usr/share/keyrings/docker-archive-keyring.gpg echo \ "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/docker-archive-keyring.gpg] https://download.docker.com/linux/ubuntu \ $(lsb_release -cs) stable" | sudo tee /etc/apt/sources.list.d/docker.list > /dev/null sudo apt-get update sudo apt-get -y autoremove sudo apt-get -y autoclean sudo apt-get -y install docker-ce docker-ce-cli containerd.io sudo usermod -aG docker sshuser sudo service docker restart sudo chmod 666 /var/run/docker.sock # Verify that we have a working docker setup now. docker run hello-world
Configure Yarn for running Docker workloads
Run following python script as $ sudo python configure_docker_on_yarn.py
Copying the following formatted script ends up giving some whitespaces in container-executor.cfg
. Before running the python script below, please make sure to remove whitespaces from those empty lines before section headers [docker]
, [gpu]
and [cgroups]
in the script; otherwise container-executor
will complain about invalid configuration
# /usr/bin/python # configure_docker_on_yarn.py import sys import subprocess from datetime import datetime from hdinsight_common.AmbariHelper import AmbariHelper ambari_helper = AmbariHelper() current_ts=datetime.today().strftime('%Y_%m_%d_%H_%M_%S') hdp_version=subprocess.check_output(["/usr/bin/hdp-select", "status", "spark3-client"]).split("-")[2].strip() new_yarn_config = { "yarn.nodemanager.container-executor.class": "org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor", "yarn.nodemanager.linux-container-executor.group": "hadoop", "yarn.nodemanager.linux-container-executor.nonsecure-mode.limit-users": "false", "yarn.nodemanager.runtime.linux.allowed-runtimes": "default,docker", "yarn.nodemanager.runtime.linux.docker.allowed-container-networks": "host,none,bridge", "yarn.nodemanager.runtime.linux.docker.default-container-network": "host", "yarn.nodemanager.runtime.linux.docker.host-pid-namespace.allowed": "false", "yarn.nodemanager.runtime.linux.docker.privileged-containers.allowed": "false", "yarn.nodemanager.runtime.linux.docker.privileged-containers.acl": "", "yarn.nodemanager.runtime.linux.docker.capabilities": "CHOWN,DAC_OVERRIDE,FSETID,FOWNER,MKNOD,NET_RAW,SETGID,SETUID,SETFCAP,SETPCAP,NET_BIND_SERVICE,SYS_CHROOT,KILL,AUDIT_WRITE", "yarn.nodemanager.runtime.linux.docker.delayed-removal.allowed": "true", "yarn.nodemanager.delete.debug-delay-sec": "900" } # Apply the configurations to yarn-site ambari_helper.update_latest_service_config("yarn-site", "YARN", "update_yarn_site_for_docker_" + current_ts, new_yarn_config) container_executor_cfg=""" yarn.nodemanager.local-dirs=/mnt/resource/hadoop/yarn/local yarn.nodemanager.log-dirs=/mnt/resource/hadoop/yarn/log yarn.nodemanager.linux-container-executor.group=hadoop banned.users=hdfs,yarn,mapred,bin min.user.id=1000 [docker] module.enabled=true docker.binary=/usr/bin/docker docker.allowed.capabilities=CHOWN,DAC_OVERRIDE,FSETID,FOWNER,MKNOD,NET_RAW,SETGID,SETUID,SETFCAP,SETPCAP,NET_BIND_SERVICE,SYS_CHROOT,KILL,AUDIT_WRITE docker.allowed.devices= docker.allowed.networks=host,none,bridge docker.allowed.ro-mounts=/etc/passwd,/etc/group,/etc/hadoop/conf,/mnt/resource/hadoop/yarn/local,/usr/lib/hdinsight-common,/usr/lib/hdinsight-spark,/usr/lib/hdinsight-logging,/usr/hdp/{}/spark3/jars docker.allowed.rw-mounts=/mnt/resource/hadoop/yarn/local,/mnt/resource/hadoop/yarn/log docker.privileged-containers.enabled=false docker.trusted.registries=local docker.allowed.volume-drivers= [gpu] module.enabled=false [cgroups] root=/sys/fs/cgroup yarn-hierarchy=yarn """.format(hdp_version) container_executor_properties = { "content": container_executor_cfg } # Update container-executor.cfg ambari_helper.update_latest_service_config("container-executor", "YARN", "update_container_executor_configs_" + current_ts, container_executor_properties) # Restart all services so the new configurations take effect. request_id=ambari_helper.restart_all_stale_services() if request_id: ambari_helper.wait_for_request_completion(request_id, 600, 60) else: print("Failed to restart all stale services")
Verify that container-executor.cfg
is looking good
HDP_VERSION=$(hdp-select status hadoop-yarn-nodemanager | awk '{print $3}') sudo -u yarn /usr/hdp/${HDP_VERSION}/hadoop-yarn/bin/container-executor --checksetup && echo "Successful!"
Clean up usercache
directory on the workernode
This is required so when container-executor
runs, it can setup the directory with appropriate permissions that would allow docker containers to mount local paths in ro
and rw
modes as necessary. This is configured in /etc/hadoop/conf/yarn-site.xml
in yarn.nodemanager.local-dirs
and defaults to /mnt/resource/hadoop/yarn/local
. Example command if the default value is used:
NODEMANAGER_LOCAL_DIR=/mnt/resource/hadoop/yarn/local sudo rm -rf ${NODEMANAGER_LOCAL_DIR}/usercache/spark
Set up a container image locally on the worker node
We will use the following as our Dockerfile
FROM adoptopenjdk/openjdk8 RUN apt-get update -qq > /dev/null \ # Required for hadoop && apt install -y libsnappy-dev \ # Useful tools for debugging/development && apt install -y wget vim telnet lsof # Python 2 is needed for some utility scripts on HDInsight clusters RUN apt install python python-dev -y \ && curl https://bootstrap.pypa.io/pip/2.7/get-pip.py -o get-pip.py \ && python get-pip.py \ && python -V \ && pip install virtualenv \ && pip install --upgrade pip RUN apt install python3 python3-dev python3-pip python3-virtualenv -y \ && python3 -V \ && pip3 install --upgrade pip # Default to python3 since python2 has been completely removed from latest Ubuntu release and is not recommended for any new development ENV PYSPARK_DRIVER_PYTHON python3 ENV PYSPARK_PYTHON python3 RUN pip3 install numpy pandas \ && python3 -c "import numpy as np"
Now build a container image that we are going to call local/pyspark
, in production this could be coming from other repositories that are specifically configured and allowed
in container-executor.cfg
(see above).
$ ls -l pyspark/ total 4 -rw-rw-r-- 1 sshuser sshuser 478 Feb 18 02:03 Dockerfile $ docker build -t local/pyspark pyspark/
Verify that local image works
$ docker run -it local/pyspark bash -c "(java -version && python -V && python3 -V)"
Run pi
spark job on the cluster
Run following:
PYSPARK_IMAGE=local/pyspark SPARK_EXAMPLES_JAR=/usr/hdp/current/spark3-client/examples/jars/spark-examples.jar HDP_VERSION=$(hdp-select status spark3-client | awk '{print $3}') SPARK_JARS_PATH=/usr/hdp/${HDP_VERSION}/spark3/jars SPARK_CUSTOM_JARS_PATH=/usr/lib/hdinsight-spark HDI_LOGGING=/usr/lib/hdinsight-logging HDI_COMMON=/usr/lib/hdinsight-common HADOOP_CONF_DIR=/etc/hadoop/conf CUSTOM_CLASSPATH="${SPARK_JARS_PATH}/*:${SPARK_CUSTOM_JARS_PATH}/*:${HDI_LOGGING}/*" MOUNTS="/etc/passwd:/etc/passwd:ro,/etc/group:/etc/group:ro,${HADOOP_CONF_DIR}:${HADOOP_CONF_DIR}:ro,${HDI_COMMON}:${HDI_COMMON}:ro,${HDI_LOGGING}:${HDI_LOGGING}:ro,${SPARK_JARS_PATH}:${SPARK_JARS_PATH}:ro,${SPARK_CUSTOM_JARS_PATH}:${SPARK_CUSTOM_JARS_PATH}:ro" spark-submit --class org.apache.spark.examples.SparkPi \ --master yarn --deploy-mode cluster --num-executors 1 \ --driver-memory 2g --executor-memory 1g --executor-cores 2 \ --conf "spark.driver.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf "spark.executor.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf "spark.yarn.am.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_TYPE=docker \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=${PYSPARK_IMAGE} \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_MOUNTS=${MOUNTS} \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_DELAYED_REMOVAL=true \ --conf spark.executor.extraClassPath="${CUSTOM_CLASSPATH}" \ --conf spark.driver.extraClassPath="${CUSTOM_CLASSPATH}" \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_TYPE=docker \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=${PYSPARK_IMAGE} \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_MOUNTS=${MOUNTS} \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_DELAYED_REMOVAL=true \ ${SPARK_EXAMPLES_JAR} 10
In the end we should see something like below:
... ... 23/02/19 03:36:44 INFO Client [main]: Application report for application_1676773095123_0019 (state: RUNNING) 23/02/19 03:36:45 INFO Client [main]: Application report for application_1676773095123_0019 (state: RUNNING) 23/02/19 03:36:46 INFO Client [main]: Application report for application_1676773095123_0019 (state: RUNNING) 23/02/19 03:36:47 INFO Client [main]: Application report for application_1676773095123_0019 (state: RUNNING) 23/02/19 03:36:48 INFO Client [main]: Application report for application_1676773095123_0019 (state: RUNNING) 23/02/19 03:36:49 INFO Client [main]: Application report for application_1676773095123_0019 (state: RUNNING) 23/02/19 03:36:50 INFO Client [main]: Application report for application_1676773095123_0019 (state: RUNNING) 23/02/19 03:36:51 INFO Client [main]: Application report for application_1676773095123_0019 (state: FINISHED) ...
The output is available in stdout for the application and can be seen on Yarn UI by going to https://[clustername].azurehdinsight.net/yarnui (example below):
Log Type: stdout Log Upload Time: Sun Feb 19 03:36:52 +0000 2023 Log Length: 33 Pi is roughly 3.1445191445191445
Next we are going to run a simple numpy job. We already installed numpy on our local/pyspark
container image, so we don't need anything additional here. First create the sample script:
# numpy_example.py from pyspark.sql import SparkSession spark = SparkSession.builder.appName("numpy_example").getOrCreate() sc = spark.sparkContext import numpy as np from numpy import random arr = np.arange(15).reshape(3, 5) print("Original array: ", arr) random.shuffle(arr) print("Array after shuffling: ", arr)
Now run the following (output will again be available on stdout for the application on Yarn UI):
PYSPARK_IMAGE=local/pyspark SPARK_EXAMPLES_JAR=/usr/hdp/current/spark3-client/examples/jars/spark-examples.jar HDP_VERSION=$(hdp-select status spark3-client | awk '{print $3}') SPARK_JARS_PATH=/usr/hdp/${HDP_VERSION}/spark3/jars SPARK_CUSTOM_JARS_PATH=/usr/lib/hdinsight-spark HDI_LOGGING=/usr/lib/hdinsight-logging HDI_COMMON=/usr/lib/hdinsight-common HADOOP_CONF_DIR=/etc/hadoop/conf CUSTOM_CLASSPATH="${SPARK_JARS_PATH}/*:${SPARK_CUSTOM_JARS_PATH}/*:${HDI_LOGGING}/*" MOUNTS="/etc/passwd:/etc/passwd:ro,/etc/group:/etc/group:ro,${HADOOP_CONF_DIR}:${HADOOP_CONF_DIR}:ro,${HDI_COMMON}:${HDI_COMMON}:ro,${HDI_LOGGING}:${HDI_LOGGING}:ro,${SPARK_JARS_PATH}:${SPARK_JARS_PATH}:ro,${SPARK_CUSTOM_JARS_PATH}:${SPARK_CUSTOM_JARS_PATH}:ro" spark-submit --master yarn --deploy-mode cluster --num-executors 1 \ --driver-memory 2g --executor-memory 1g --executor-cores 2 \ --conf "spark.driver.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf "spark.executor.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf "spark.yarn.am.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_TYPE=docker \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=${PYSPARK_IMAGE} \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_MOUNTS=${MOUNTS} \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_DELAYED_REMOVAL=true \ --conf spark.executor.extraClassPath="${CUSTOM_CLASSPATH}" \ --conf spark.driver.extraClassPath="${CUSTOM_CLASSPATH}" \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_TYPE=docker \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=${PYSPARK_IMAGE} \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_MOUNTS=${MOUNTS} \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_DELAYED_REMOVAL=true \ numpy_example.py
The output should look something like below in stdout on YarnUI:
Log Type: stdout Log Upload Time: Sun Feb 19 04:37:08 +0000 2023 Log Length: 151 Original array: [[ 0 1 2 3 4] [ 5 6 7 8 9] [10 11 12 13 14]] Array after shuffling: [[ 0 1 2 3 4] [10 11 12 13 14] [ 5 6 7 8 9]]
Similar command to launch spark-shell
will be:
PYSPARK_IMAGE=local/pyspark SPARK_EXAMPLES_JAR=/usr/hdp/current/spark3-client/examples/jars/spark-examples.jar HDP_VERSION=$(hdp-select status spark3-client | awk '{print $3}') SPARK_JARS_PATH=/usr/hdp/${HDP_VERSION}/spark3/jars SPARK_CUSTOM_JARS_PATH=/usr/lib/hdinsight-spark HDI_LOGGING=/usr/lib/hdinsight-logging HDI_COMMON=/usr/lib/hdinsight-common HADOOP_CONF_DIR=/etc/hadoop/conf CUSTOM_CLASSPATH="${SPARK_JARS_PATH}/*:${SPARK_CUSTOM_JARS_PATH}/*:${HDI_LOGGING}/*" MOUNTS="/etc/passwd:/etc/passwd:ro,/etc/group:/etc/group:ro,${HADOOP_CONF_DIR}:${HADOOP_CONF_DIR}:ro,${HDI_COMMON}:${HDI_COMMON}:ro,${HDI_LOGGING}:${HDI_LOGGING}:ro,${SPARK_JARS_PATH}:${SPARK_JARS_PATH}:ro,${SPARK_CUSTOM_JARS_PATH}:${SPARK_CUSTOM_JARS_PATH}:ro" spark-shell --master yarn --num-executors 1 \ --driver-memory 2g --executor-memory 1g --executor-cores 2 \ --conf "spark.driver.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf "spark.executor.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf "spark.yarn.am.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_TYPE=docker \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=${PYSPARK_IMAGE} \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_MOUNTS=${MOUNTS} \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_DELAYED_REMOVAL=true \ --conf spark.executor.extraClassPath="${CUSTOM_CLASSPATH}" \ --conf spark.driver.extraClassPath="${CUSTOM_CLASSPATH}" \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_TYPE=docker \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=${PYSPARK_IMAGE} \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_MOUNTS=${MOUNTS} \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_DELAYED_REMOVAL=true
And for pyspark
will be:
PYSPARK_IMAGE=local/pyspark SPARK_EXAMPLES_JAR=/usr/hdp/current/spark3-client/examples/jars/spark-examples.jar HDP_VERSION=$(hdp-select status spark3-client | awk '{print $3}') SPARK_JARS_PATH=/usr/hdp/${HDP_VERSION}/spark3/jars SPARK_CUSTOM_JARS_PATH=/usr/lib/hdinsight-spark HDI_LOGGING=/usr/lib/hdinsight-logging HDI_COMMON=/usr/lib/hdinsight-common HADOOP_CONF_DIR=/etc/hadoop/conf CUSTOM_CLASSPATH="${SPARK_JARS_PATH}/*:${SPARK_CUSTOM_JARS_PATH}/*:${HDI_LOGGING}/*" MOUNTS="/etc/passwd:/etc/passwd:ro,/etc/group:/etc/group:ro,${HADOOP_CONF_DIR}:${HADOOP_CONF_DIR}:ro,${HDI_COMMON}:${HDI_COMMON}:ro,${HDI_LOGGING}:${HDI_LOGGING}:ro,${SPARK_JARS_PATH}:${SPARK_JARS_PATH}:ro,${SPARK_CUSTOM_JARS_PATH}:${SPARK_CUSTOM_JARS_PATH}:ro" pyspark --master yarn --num-executors 1 \ --driver-memory 2g --executor-memory 1g --executor-cores 2 \ --conf "spark.driver.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf "spark.executor.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf "spark.yarn.am.extraJavaOptions=-Dhdp.version=${HDP_VERSION}" \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_TYPE=docker \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=${PYSPARK_IMAGE} \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_MOUNTS=${MOUNTS} \ --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_DELAYED_REMOVAL=true \ --conf spark.executor.extraClassPath="${CUSTOM_CLASSPATH}" \ --conf spark.driver.extraClassPath="${CUSTOM_CLASSPATH}" \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_TYPE=docker \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=${PYSPARK_IMAGE} \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_MOUNTS=${MOUNTS} \ --conf spark.yarn.appMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_DELAYED_REMOVAL=true
In this post, we ran Spark pi
job and a simple numpy
script on docker containers on yarn. We also ran spark-shell
and pyspark
that would run driver and executors on docker containers. Running the shells can be very helpful in debugging container issues because the containers don't exit even on job failure and that means we can run $ docker exec -it [container_name] bash
to attach a shell to the running container.