Running Spark on YARN

• Security
• Launching Spark on YARN
  • Adding Other JARs
• Preparations
• Configuration
• Debugging your Application
  • Spark Properties
  • Available patterns for SHS custom executor log URL
• Resource Allocation and Configuration Overview
• Stage Level Scheduling Overview
• Important notes
• Kerberos
  • YARN-specific Kerberos Configuration
  • Troubleshooting Kerberos
• Configuring the External Shuffle Service
• Launching your application with Apache Oozie
• Using the Spark History Server to replace the Spark Web UI
• Running multiple versions of the Spark Shuffle Service

Support for running on YARN (Hadoop NextGen) was added to Spark in version 0.6.0, and improved in subsequent releases.

Security

Security features like authentication are not enabled by default. When deploying a cluster that is open to the internet or an untrusted network, it’s important to secure access to the cluster to prevent unauthorized applications from running on the cluster. Please see Spark Security and the specific security sections in this doc before running Spark.

Launching Spark on YARN

Ensure that HADOOP_CONF_DIR or YARN_CONF_DIR points to the directory which contains the (client side) configuration files for the Hadoop cluster. These configs are used to write to HDFS and connect to the YARN ResourceManager. The configuration contained in this directory will be distributed to the YARN cluster so that all containers used by the application use the same configuration. If the configuration references Java system properties or environment variables not managed by YARN, they should also be set in the Spark application’s configuration (driver, executors, and the AM when running in client mode).

There are two deploy modes that can be used to launch Spark applications on YARN. In cluster mode, the Spark driver runs inside an application master process which is managed by YARN on the cluster, and the client can go away after initiating the application. In client mode, the driver runs in the client process, and the application master is only used for requesting resources from YARN.

Unlike other cluster managers supported by Spark in which the master’s address is specified in the --master parameter, in YARN mode the ResourceManager’s address is picked up from the Hadoop configuration. Thus, the --master parameter is yarn.

To launch a Spark application in cluster mode:

$ ./bin/spark-submit --class path.to.your.Class --master yarn --deploy-mode cluster [options] <app jar> [app options]

For example:

$ ./bin/spark-submit --class org.apache.spark.examples.SparkPi \
    --master yarn \
    --deploy-mode cluster \
    --driver-memory 4g \
    --executor-memory 2g \
    --executor-cores 1 \
    --queue thequeue \
    examples/jars/spark-examples*.jar \
    10

The above starts a YARN client program which starts the default Application Master. Then SparkPi will be run as a child thread of Application Master. The client will periodically poll the Application Master for status updates and display them in the console. The client will exit once your application has finished running. Refer to the Debugging your Application section below for how to see driver and executor logs.

To launch a Spark application in client mode, do the same, but replace cluster with client. The following shows how you can run spark-shell in client mode:

$ ./bin/spark-shell --master yarn --deploy-mode client

Adding Other JARs

In cluster mode, the driver runs on a different machine than the client, so SparkContext.addJar won’t work out of the box with files that are local to the client. To make files on the client available to SparkContext.addJar, include them with the --jars option in the launch command.

$ ./bin/spark-submit --class my.main.Class \
    --master yarn \
    --deploy-mode cluster \
    --jars my-other-jar.jar,my-other-other-jar.jar \
    my-main-jar.jar \
    app_arg1 app_arg2

Preparations

Running Spark on YARN requires a binary distribution of Spark which is built with YARN support. Binary distributions can be downloaded from the downloads page of the project website. There are two variants of Spark binary distributions you can download. One is pre-built with a certain version of Apache Hadoop; this Spark distribution contains built-in Hadoop runtime, so we call it with-hadoop Spark distribution. The other one is pre-built with user-provided Hadoop; since this Spark distribution doesn’t contain a built-in Hadoop runtime, it’s smaller, but users have to provide a Hadoop installation separately. We call this variant no-hadoop Spark distribution. For with-hadoop Spark distribution, since it contains a built-in Hadoop runtime already, by default, when a job is submitted to Hadoop Yarn cluster, to prevent jar conflict, it will not populate Yarn’s classpath into Spark. To override this behavior, you can set spark.yarn.populateHadoopClasspath=true. For no-hadoop Spark distribution, Spark will populate Yarn’s classpath by default in order to get Hadoop runtime. For with-hadoop Spark distribution, if your application depends on certain library that is only available in the cluster, you can try to populate the Yarn classpath by setting the property mentioned above. If you run into jar conflict issue by doing so, you will need to turn it off and include this library in your application jar.

To build Spark yourself, refer to Building Spark.

To make Spark runtime jars accessible from YARN side, you can specify spark.yarn.archive or spark.yarn.jars. For details please refer to Spark Properties. If neither spark.yarn.archive nor spark.yarn.jars is specified, Spark will create a zip file with all jars under $SPARK_HOME/jars and upload it to the distributed cache.

Configuration

Most of the configs are the same for Spark on YARN as for other deployment modes. See the configuration page for more information on those. These are configs that are specific to Spark on YARN.

Debugging your Application

In YARN terminology, executors and application masters run inside “containers”. YARN has two modes for handling container logs after an application has completed. If log aggregation is turned on (with the yarn.log-aggregation-enable config), container logs are copied to HDFS and deleted on the local machine. These logs can be viewed from anywhere on the cluster with the yarn logs command.

yarn logs -applicationId <app ID>

will print out the contents of all log files from all containers from the given application. You can also view the container log files directly in HDFS using the HDFS shell or API. The directory where they are located can be found by looking at your YARN configs (yarn.nodemanager.remote-app-log-dir and yarn.nodemanager.remote-app-log-dir-suffix). The logs are also available on the Spark Web UI under the Executors Tab. You need to have both the Spark history server and the MapReduce history server running and configure yarn.log.server.url in yarn-site.xml properly. The log URL on the Spark history server UI will redirect you to the MapReduce history server to show the aggregated logs.

When log aggregation isn’t turned on, logs are retained locally on each machine under YARN_APP_LOGS_DIR, which is usually configured to /tmp/logs or $HADOOP_HOME/logs/userlogs depending on the Hadoop version and installation. Viewing logs for a container requires going to the host that contains them and looking in this directory. Subdirectories organize log files by application ID and container ID. The logs are also available on the Spark Web UI under the Executors Tab and doesn’t require running the MapReduce history server.

To review per-container launch environment, increase yarn.nodemanager.delete.debug-delay-sec to a large value (e.g. 36000), and then access the application cache through yarn.nodemanager.local-dirs on the nodes on which containers are launched. This directory contains the launch script, JARs, and all environment variables used for launching each container. This process is useful for debugging classpath problems in particular. (Note that enabling this requires admin privileges on cluster settings and a restart of all node managers. Thus, this is not applicable to hosted clusters).

To use a custom log4j configuration for the application master or executors, here are the options:

• upload a custom log4j.properties using spark-submit, by adding it to the --files list of files to be uploaded with the application.
• add -Dlog4j.configuration=<location of configuration file> to spark.driver.extraJavaOptions (for the driver) or spark.executor.extraJavaOptions (for executors). Note that if using a file, the file: protocol should be explicitly provided, and the file needs to exist locally on all the nodes.
• update the $SPARK_CONF_DIR/log4j.properties file and it will be automatically uploaded along with the other configurations. Note that other 2 options has higher priority than this option if multiple options are specified.

Note that for the first option, both executors and the application master will share the same log4j configuration, which may cause issues when they run on the same node (e.g. trying to write to the same log file).

If you need a reference to the proper location to put log files in the YARN so that YARN can properly display and aggregate them, use spark.yarn.app.container.log.dir in your log4j.properties. For example, log4j.appender.file_appender.File=${spark.yarn.app.container.log.dir}/spark.log. For streaming applications, configuring RollingFileAppender and setting file location to YARN’s log directory will avoid disk overflow caused by large log files, and logs can be accessed using YARN’s log utility.

To use a custom metrics.properties for the application master and executors, update the $SPARK_CONF_DIR/metrics.properties file. It will automatically be uploaded with other configurations, so you don’t need to specify it manually with --files.

Spark Properties

Available patterns for SHS custom executor log URL

For example, suppose you would like to point log url link to Job History Server directly instead of let NodeManager http server redirects it, you can configure spark.history.custom.executor.log.url as below:

{{HTTP_SCHEME}}<JHS_HOST>:<JHS_PORT>/jobhistory/logs/{{NM_HOST}}:{{NM_PORT}}/{{CONTAINER_ID}}/{{CONTAINER_ID}}/{{USER}}/{{FILE_NAME}}?start=-4096

NOTE: you need to replace <JHS_HOST> and <JHS_PORT> with actual value.

Resource Allocation and Configuration Overview

Please make sure to have read the Custom Resource Scheduling and Configuration Overview section on the configuration page. This section only talks about the YARN specific aspects of resource scheduling.

YARN needs to be configured to support any resources the user wants to use with Spark. Resource scheduling on YARN was added in YARN 3.1.0. See the YARN documentation for more information on configuring resources and properly setting up isolation. Ideally the resources are setup isolated so that an executor can only see the resources it was allocated. If you do not have isolation enabled, the user is responsible for creating a discovery script that ensures the resource is not shared between executors.

YARN supports user defined resource types but has built in types for GPU (yarn.io/gpu) and FPGA (yarn.io/fpga). For that reason, if you are using either of those resources, Spark can translate your request for spark resources into YARN resources and you only have to specify the spark.{driver/executor}.resource. configs. Note, if you are using a custom resource type for GPUs or FPGAs with YARN you can change the Spark mapping using spark.yarn.resourceGpuDeviceName and spark.yarn.resourceFpgaDeviceName. If you are using a resource other than FPGA or GPU, the user is responsible for specifying the configs for both YARN (spark.yarn.{driver/executor}.resource.) and Spark (spark.{driver/executor}.resource.).

For example, the user wants to request 2 GPUs for each executor. The user can just specify spark.executor.resource.gpu.amount=2 and Spark will handle requesting yarn.io/gpu resource type from YARN.

If the user has a user defined YARN resource, lets call it acceleratorX then the user must specify spark.yarn.executor.resource.acceleratorX.amount=2 and spark.executor.resource.acceleratorX.amount=2.

YARN does not tell Spark the addresses of the resources allocated to each container. For that reason, the user must specify a discovery script that gets run by the executor on startup to discover what resources are available to that executor. You can find an example scripts in examples/src/main/scripts/getGpusResources.sh. The script must have execute permissions set and the user should setup permissions to not allow malicious users to modify it. The script should write to STDOUT a JSON string in the format of the ResourceInformation class. This has the resource name and an array of resource addresses available to just that executor.

Stage Level Scheduling Overview

Stage level scheduling is supported on YARN when dynamic allocation is enabled. One thing to note that is YARN specific is that each ResourceProfile requires a different container priority on YARN. The mapping is simply the ResourceProfile id becomes the priority, on YARN lower numbers are higher priority. This means that profiles created earlier will have a higher priority in YARN. Normally this won’t matter as Spark finishes one stage before starting another one, the only case this might have an affect is in a job server type scenario, so its something to keep in mind. Note there is a difference in the way custom resources are handled between the base default profile and custom ResourceProfiles. To allow for the user to request YARN containers with extra resources without Spark scheduling on them, the user can specify resources via the spark.yarn.executor.resource. config. Those configs are only used in the base default profile though and do not get propagated into any other custom ResourceProfiles. This is because there would be no way to remove them if you wanted a stage to not have them. This results in your default profile getting custom resources defined in spark.yarn.executor.resource. plus spark defined resources of GPU or FPGA. Spark converts GPU and FPGA resources into the YARN built in types yarn.io/gpu) and yarn.io/fpga, but does not know the mapping of any other resources. Any other Spark custom resources are not propagated to YARN for the default profile. So if you want Spark to schedule based off a custom resource and have it requested from YARN, you must specify it in both YARN (spark.yarn.{driver/executor}.resource.) and Spark (spark.{driver/executor}.resource.) configs. Leave the Spark config off if you only want YARN containers with the extra resources but Spark not to schedule using them. Now for custom ResourceProfiles, it doesn’t currently have a way to only specify YARN resources without Spark scheduling off of them. This means for custom ResourceProfiles we propagate all the resources defined in the ResourceProfile to YARN. We still convert GPU and FPGA to the YARN build in types as well. This requires that the name of any custom resources you specify match what they are defined as in YARN.

Important notes

• Whether core requests are honored in scheduling decisions depends on which scheduler is in use and how it is configured.
• In cluster mode, the local directories used by the Spark executors and the Spark driver will be the local directories configured for YARN (Hadoop YARN config yarn.nodemanager.local-dirs). If the user specifies spark.local.dir, it will be ignored. In client mode, the Spark executors will use the local directories configured for YARN while the Spark driver will use those defined in spark.local.dir. This is because the Spark driver does not run on the YARN cluster in client mode, only the Spark executors do.
• The --files and --archives options support specifying file names with the # similar to Hadoop. For example, you can specify: --files localtest.txt#appSees.txt and this will upload the file you have locally named localtest.txt into HDFS but this will be linked to by the name appSees.txt, and your application should use the name as appSees.txt to reference it when running on YARN.
• The --jars option allows the SparkContext.addJar function to work if you are using it with local files and running in cluster mode. It does not need to be used if you are using it with HDFS, HTTP, HTTPS, or FTP files.

Kerberos

Standard Kerberos support in Spark is covered in the Security page.

In YARN mode, when accessing Hadoop file systems, aside from the default file system in the hadoop configuration, Spark will also automatically obtain delegation tokens for the service hosting the staging directory of the Spark application.

YARN-specific Kerberos Configuration

Troubleshooting Kerberos

Debugging Hadoop/Kerberos problems can be “difficult”. One useful technique is to enable extra logging of Kerberos operations in Hadoop by setting the HADOOP_JAAS_DEBUG environment variable.

export HADOOP_JAAS_DEBUG=true

The JDK classes can be configured to enable extra logging of their Kerberos and SPNEGO/REST authentication via the system properties sun.security.krb5.debug and sun.security.spnego.debug=true

-Dsun.security.krb5.debug=true -Dsun.security.spnego.debug=true

All these options can be enabled in the Application Master:

spark.yarn.appMasterEnv.HADOOP_JAAS_DEBUG true
spark.yarn.am.extraJavaOptions -Dsun.security.krb5.debug=true -Dsun.security.spnego.debug=true

Finally, if the log level for org.apache.spark.deploy.yarn.Client is set to DEBUG, the log will include a list of all tokens obtained, and their expiry details

Configuring the External Shuffle Service

To start the Spark Shuffle Service on each NodeManager in your YARN cluster, follow these instructions:

1. Build Spark with the YARN profile. Skip this step if you are using a pre-packaged distribution.
2. Locate the spark-<version>-yarn-shuffle.jar. This should be under $SPARK_HOME/common/network-yarn/target/scala-<version> if you are building Spark yourself, and under yarn if you are using a distribution.
3. Add this jar to the classpath of all NodeManagers in your cluster.
4. In the yarn-site.xml on each node, add spark_shuffle to yarn.nodemanager.aux-services, then set yarn.nodemanager.aux-services.spark_shuffle.class to org.apache.spark.network.yarn.YarnShuffleService.
5. Increase NodeManager's heap size by setting YARN_HEAPSIZE (1000 by default) in etc/hadoop/yarn-env.sh to avoid garbage collection issues during shuffle.
6. Restart all NodeManagers in your cluster.

The following extra configuration options are available when the shuffle service is running on YARN:

Please note that the instructions above assume that the default shuffle service name, spark_shuffle, has been used. It is possible to use any name here, but the values used in the YARN NodeManager configurations must match the value of spark.shuffle.service.name in the Spark application.

The shuffle service will, by default, take all of its configurations from the Hadoop Configuration used by the NodeManager (e.g. yarn-site.xml). However, it is also possible to configure the shuffle service independently using a file named spark-shuffle-site.xml which should be placed onto the classpath of the shuffle service (which is, by default, shared with the classpath of the NodeManager). The shuffle service will treat this as a standard Hadoop Configuration resource and overlay it on top of the NodeManager’s configuration.

Launching your application with Apache Oozie

Apache Oozie can launch Spark applications as part of a workflow. In a secure cluster, the launched application will need the relevant tokens to access the cluster’s services. If Spark is launched with a keytab, this is automatic. However, if Spark is to be launched without a keytab, the responsibility for setting up security must be handed over to Oozie.

The details of configuring Oozie for secure clusters and obtaining credentials for a job can be found on the Oozie web site in the “Authentication” section of the specific release’s documentation.

For Spark applications, the Oozie workflow must be set up for Oozie to request all tokens which the application needs, including:

• The YARN resource manager.
• The local Hadoop filesystem.
• Any remote Hadoop filesystems used as a source or destination of I/O.
• Hive —if used.
• HBase —if used.
• The YARN timeline server, if the application interacts with this.

To avoid Spark attempting —and then failing— to obtain Hive, HBase and remote HDFS tokens, the Spark configuration must be set to disable token collection for the services.

The Spark configuration must include the lines:

spark.security.credentials.hive.enabled   false
spark.security.credentials.hbase.enabled  false

The configuration option spark.kerberos.access.hadoopFileSystems must be unset.

Using the Spark History Server to replace the Spark Web UI

It is possible to use the Spark History Server application page as the tracking URL for running applications when the application UI is disabled. This may be desirable on secure clusters, or to reduce the memory usage of the Spark driver. To set up tracking through the Spark History Server, do the following:

• On the application side, set spark.yarn.historyServer.allowTracking=true in Spark’s configuration. This will tell Spark to use the history server’s URL as the tracking URL if the application’s UI is disabled.
• On the Spark History Server, add org.apache.spark.deploy.yarn.YarnProxyRedirectFilter to the list of filters in the spark.ui.filters configuration.

Be aware that the history server information may not be up-to-date with the application’s state.

Running multiple versions of the Spark Shuffle Service

Please note that this section only applies when running on YARN versions >= 2.9.0.

In some cases it may be desirable to run multiple instances of the Spark Shuffle Service which are using different versions of Spark. This can be helpful, for example, when running a YARN cluster with a mixed workload of applications running multiple Spark versions, since a given version of the shuffle service is not always compatible with other versions of Spark. YARN versions since 2.9.0 support the ability to run shuffle services within an isolated classloader (see YARN-4577), meaning multiple Spark versions can coexist within a single NodeManager. The yarn.nodemanager.aux-services.<service-name>.classpath and, starting from YARN 2.10.2/3.1.1/3.2.0, yarn.nodemanager.aux-services.<service-name>.remote-classpath options can be used to configure this. Note that YARN 3.3.0/3.3.1 have an issue which requires setting yarn.nodemanager.aux-services.<service-name>.system-classes as a workaround. See YARN-11053 for details. In addition to setting up separate classpaths, it’s necessary to ensure the two versions advertise to different ports. This can be achieved using the spark-shuffle-site.xml file described above. For example, you may have configuration like:

  yarn.nodemanager.aux-services = spark_shuffle_x,spark_shuffle_y
  yarn.nodemanager.aux-services.spark_shuffle_x.classpath = /path/to/spark-x-path/fat.jar:/path/to/spark-x-config
  yarn.nodemanager.aux-services.spark_shuffle_y.classpath = /path/to/spark-y-path/fat.jar:/path/to/spark-y-config

Or

  yarn.nodemanager.aux-services = spark_shuffle_x,spark_shuffle_y
  yarn.nodemanager.aux-services.spark_shuffle_x.classpath = /path/to/spark-x-path/*:/path/to/spark-x-config
  yarn.nodemanager.aux-services.spark_shuffle_y.classpath = /path/to/spark-y-path/*:/path/to/spark-y-config

The two spark-*-config directories each contain one file, spark-shuffle-site.xml. These are XML files in the Hadoop Configuration format which each contain a few configurations to adjust the port number and metrics name prefix used:

<configuration>
  <property>
    <name>spark.shuffle.service.port</name>
    <value>7001</value>
  </property>
  <property>
    <name>spark.yarn.shuffle.service.metrics.namespace</name>
    <value>sparkShuffleServiceX</value>
  </property>
</configuration>

The values should both be different for the two different services.

Then, in the configuration of the Spark applications, one should be configured with:

  spark.shuffle.service.name = spark_shuffle_x
  spark.shuffle.service.port = 7001

and one should be configured with:

  spark.shuffle.service.name = spark_shuffle_y
  spark.shuffle.service.port = <other value>