Launching Applications Using Docker Containers

Security Warning

IMPORTANT This feature is experimental and is not complete. IMPORTANT Enabling this feature and running Docker containers in your cluster has security implications. With this feature enabled, it may be possible to gain root access to the YARN NodeManager hosts. Given Docker’s integration with many powerful kernel features, it is imperative that administrators understand Docker security before enabling this feature.

Overview

Docker combines an easy-to-use interface to Linux containers with easy-to-construct image files for those containers. In short, Docker enables users to bundle an application together with its preferred execution environment to be executed on a target machine. For more information about Docker, see their documentation.

The Linux Container Executor (LCE) allows the YARN NodeManager to launch YARN containers to run either directly on the host machine or inside Docker containers. The application requesting the resources can specify for each container how it should be executed. The LCE also provides enhanced security and is required when deploying a secure cluster. When the LCE launches a YARN container to execute in a Docker container, the application can specify the Docker image to be used.

Docker containers provide a custom execution environment in which the application’s code runs, isolated from the execution environment of the NodeManager and other applications. These containers can include special libraries needed by the application, and they can have different versions of native tools and libraries including Perl, Python, and Java. Docker containers can even run a different flavor of Linux than what is running on the NodeManager.

Docker for YARN provides both consistency (all YARN containers will have the same software environment) and isolation (no interference with whatever is installed on the physical machine).

Docker support in the LCE is still evolving. To track progress, follow YARN-3611, the umbrella JIRA for Docker support improvements.

Cluster Configuration

The LCE requires that container-executor binary be owned by root:hadoop and have 6050 permissions. In order to launch Docker containers, the Docker daemon must be running on all NodeManager hosts where Docker containers will be launched. The Docker client must also be installed on all NodeManager hosts where Docker containers will be launched and able to start Docker containers.

To prevent timeouts while starting jobs, any large Docker images to be used by an application should already be loaded in the Docker daemon’s cache on the NodeManager hosts. A simple way to load an image is by issuing a Docker pull request. For example:

    sudo docker pull images/hadoop-docker:latest

The following properties should be set in yarn-site.xml:

<configuration>
  <property>
    <name>yarn.nodemanager.container-executor.class</name>
    <value>org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor</value>
    <description>
      This is the container executor setting that ensures that all applications
      are started with the LinuxContainerExecutor.
    </description>
  </property>

  <property>
    <name>yarn.nodemanager.linux-container-executor.group</name>
    <value>hadoop</value>
    <description>
      The POSIX group of the NodeManager. It should match the setting in
      "container-executor.cfg". This configuration is required for validating
      the secure access of the container-executor binary.
    </description>
  </property>

  <property>
    <name>yarn.nodemanager.linux-container-executor.nonsecure-mode.limit-users</name>
    <value>false</value>
    <description>
      Whether all applications should be run as the NodeManager process' owner.
      When false, applications are launched instead as the application owner.
    </description>
  </property>

  <property>
    <name>yarn.nodemanager.runtime.linux.allowed-runtimes</name>
    <value>default,docker</value>
    <description>
      Comma separated list of runtimes that are allowed when using
      LinuxContainerExecutor. The allowed values are default, docker, and
      javasandbox.
    </description>
  </property>

  <property>
    <name>yarn.nodemanager.runtime.linux.docker.allowed-container-networks</name>
    <value>host,none,bridge</value>
    <description>
      Optional. A comma-separated set of networks allowed when launching
      containers. Valid values are determined by Docker networks available from
      `docker network ls`
    </description>
  </property>

  <property>
    <name>yarn.nodemanager.runtime.linux.docker.default-container-network</name>
    <value>host</value>
    <description>
      The network used when launching Docker containers when no
      network is specified in the request. This network must be one of the
      (configurable) set of allowed container networks.
    </description>
  </property>

  <property>
    <name>yarn.nodemanager.runtime.linux.docker.privileged-containers.allowed</name>
    <value>false</value>
    <description>
      Optional. Whether applications are allowed to run in privileged
      containers.
    </description>
  </property>

  <property>
    <name>yarn.nodemanager.runtime.linux.docker.privileged-containers.acl</name>
    <value></value>
    <description>
      Optional. A comma-separated list of users who are allowed to request
      privileged contains if privileged containers are allowed.
    </description>
  </property>

  <property>
    <name>yarn.nodemanager.runtime.linux.docker.capabilities</name>
    <value>CHOWN,DAC_OVERRIDE,FSETID,FOWNER,MKNOD,NET_RAW,SETGID,SETUID,SETFCAP,SETPCAP,NET_BIND_SERVICE,SYS_CHROOT,KILL,AUDIT_WRITE</value>
    <description>
      Optional. This configuration setting determines the capabilities
      assigned to docker containers when they are launched. While these may not
      be case-sensitive from a docker perspective, it is best to keep these
      uppercase. To run without any capabilites, set this value to
      "none" or "NONE"
    </description>
  </property>
</configuration>

In addition, a container-executer.cfg file must exist and contain settings for the container executor. The file must be owned by root with permissions 0400. The format of the file is the standard Java properties file format, for example

`key=value`

The following properties are required to enable Docker support:

The container-executor.cfg must contain a section to determine the capabilities that containers are allowed. It contains the following properties:

Please note that if you wish to run Docker containers that require access to the YARN local directories, you must add them to the docker.allowed.rw-mounts list.

In addition, containers are not permitted to mount any parent of the container-executor.cfg directory in read-write mode.

The following properties are optional:

Part of a container-executor.cfg which allows Docker containers to be launched is below:

yarn.nodemanager.linux-container-executor.group=yarn
[docker]
  module.enabled=true
  docker.allowed.capabilities=SYS_CHROOT,MKNOD,SETFCAP,SETPCAP,FSETID,CHOWN,AUDIT_WRITE,SETGID,NET_RAW,FOWNER,SETUID,DAC_OVERRIDE,KILL,NET_BIND_SERVICE
  docker.allowed.networks=bridge,host,none
  docker.allowed.ro-mounts=/sys/fs/cgroup
  docker.allowed.rw-mounts=/var/hadoop/yarn/local-dir,/var/hadoop/yarn/log-dir

Docker Image Requirements

In order to work with YARN, there are two requirements for Docker images.

First, the Docker container will be explicitly launched with the application owner as the container user. If the application owner is not a valid user in the Docker image, the application will fail. The container user is specified by the user’s UID. If the user’s UID is different between the NodeManager host and the Docker image, the container may be launched as the wrong user or may fail to launch because the UID does not exist.

Second, the Docker image must have whatever is expected by the application in order to execute. In the case of Hadoop (MapReduce or Spark), the Docker image must contain the JRE and Hadoop libraries and have the necessary environment variables set: JAVA_HOME, HADOOP_COMMON_PATH, HADOOP_HDFS_HOME, HADOOP_MAPRED_HOME, HADOOP_YARN_HOME, and HADOOP_CONF_DIR. Note that the Java and Hadoop component versions available in the Docker image must be compatible with what’s installed on the cluster and in any other Docker images being used for other tasks of the same job. Otherwise the Hadoop components started in the Docker container may be unable to communicate with external Hadoop components.

If a Docker image has a command set, the behavior will depend on whether the YARN_CONTAINER_RUNTIME_DOCKER_RUN_OVERRIDE_DISABLE is set to true. If so, the command will be overridden when LCE launches the image with YARN’s container launch script.

If a Docker image has an entry point set, the entry point will be honored, but the default command may be overridden, as just mentioned above. Unless the entry point is something similar to sh -c or YARN_CONTAINER_RUNTIME_DOCKER_RUN_OVERRIDE_DISABLE is set to true, the net result will likely be undesirable. Because the YARN container launch script is required to correctly launch the YARN task, use of entry points is discouraged.

If an application requests a Docker image that has not already been loaded by the Docker daemon on the host where it is to execute, the Docker daemon will implicitly perform a Docker pull command. Both MapReduce and Spark assume that tasks which take more that 10 minutes to report progress have stalled, so specifying a large Docker image may cause the application to fail.

Application Submission

Before attempting to launch a Docker container, make sure that the LCE configuration is working for applications requesting regular YARN containers. If after enabling the LCE one or more NodeManagers fail to start, the cause is most likely that the ownership and/or permissions on the container-executer binary are incorrect. Check the logs to confirm.

In order to run an application in a Docker container, set the following environment variables in the application’s environment:

The first two are required. The remainder can be set as needed. While controlling the container type through environment variables is somewhat less than ideal, it allows applications with no awareness of YARN’s Docker support (such as MapReduce and Spark) to nonetheless take advantage of it through their support for configuring the application environment.

Once an application has been submitted to be launched in a Docker container, the application will behave exactly as any other YARN application. Logs will be aggregated and stored in the relevant history server. The application life cycle will be the same as for a non-Docker application.

Connecting to a Secure Docker Repository

Until YARN-5428 is complete, the Docker client command will draw its configuration from the default location, which is $HOME/.docker/config.json on the NodeManager host. The Docker configuration is where secure repository credentials are stored, so use of the LCE with secure Docker repos is discouraged until YARN-5428 is complete.

As a work-around, you may manually log the Docker daemon on every NodeManager host into the secure repo using the Docker login command:

  docker login [OPTIONS] [SERVER]

  Register or log in to a Docker registry server, if no server is specified
  "https://index.docker.io/v1/" is the default.

  -e, --email=""       Email
  -p, --password=""    Password
  -u, --username=""    Username

Note that this approach means that all users will have access to the secure repo.

Example: MapReduce

To submit the pi job to run in Docker containers, run the following commands:

    vars="YARN_CONTAINER_RUNTIME_TYPE=docker,YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=hadoop-docker"
    hadoop jar hadoop-examples.jar pi -Dyarn.app.mapreduce.am.env=$vars \
        -Dmapreduce.map.env=$vars -Dmapreduce.reduce.env=$vars 10 100

Note that the application master, map tasks, and reduce tasks are configured independently. In this example, we are using the hadoop-docker image for all three.

Example: Spark

To run a Spark shell in Docker containers, run the following command:

    spark-shell --master yarn --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_TYPE=docker \
        --conf spark.executorEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=hadoop-docker \
        --conf spark.yarn.AppMasterEnv.YARN_CONTAINER_RUNTIME_DOCKER_IMAGE=hadoop-docker \
        --conf spark.yarn.AppMasterEnv.YARN_CONTAINER_RUNTIME_TYPE=docker

Note that the application master and executors are configured independently. In this example, we are using the hadoop-docker image for both.