Recommendations for managing Docker containers on YARN

YARN expects that Docker is already installed on all NodeManager hosts where Docker containers will run. Consider these recommendations before installing and configuring Docker for use with YARN.

Docker Version

The minimum recommended version is Docker 1.12.5. Docker is rapidly evolving and shipping multiple releases per year. Not all versions of Docker have been tested. Docker versioning changed in 2017, and is now known as Docker CE. Cloudera recommends running a recent version of Docker CE.

Note that recent versions of Docker CE use the overlay2 storage driver which probably does not work for all workloads.

RHEL/CentOs provides a version of Docker that can be installed through yum.

Storage Driver

The storage driver you choose depends on OS kernel, workload, and Docker version. Cloudera recommends administrators to read the documentation, consult with their operating system vendor, and test the desired workload before making a decision.

Tests showed that device mapper using LVM is generally stable. Under high write load to the container’s root filesystem, device mapper exhibited panics. SSDs for the Docker graph storage are recommended in this case, but care still needs to be taken. Overlay and overlay2 perform significantly better than device mapper and are recommended if the OS kernel and workload support it.

CGroup Support

YARN provides isolation through the use of cgroups. Docker also has cgroup management built in. If isolation through cgroups is desired, Cloudera recommends to use the cgroup management of YARN. YARN creates the cgroup hierarchy and set the the --cgroup-parent flag when launching the container.

The cgroupdriver must be set to cgroupfs. Ensure that Docker is running using the --exec-opt native.cgroupdriver=cgroupfs docker daemon option.

Networking

YARN has support for running Docker containers on a user specified network, however, it does not manage the Docker networks. Administrators have to create the networks before running the containers. Node labels can be used to isolate particular networks. It is crucial to read and understand the Docker networking documentation. Cloudera does not recommend swarm based options. Overlay networks can be used if the setup uses an external store, such as etcd.

YARN request the networking details, such as IP address and hostname, from the Docker. As a result, all networking types are supported. Set the YARN_CONTAINER_RUNTIME_DOCKER_CONTAINER_NETWORK environment variable to specify the network that is used.

Cloudera recommends to use host networking only for testing. If the network where the NodeManagers are running has a sufficient number of IP addresses, the bridge networking with --fixed-cidr option works properly. Each NodeManager is allocated with a small portion of the larger IP space, and then the NodeManagers allocate those IP addresses to containers.

If you want to use an administrator defined network, add the network to the yarn.nodemanager.runtime.linux.docker.allowed-container-networks property using Cloudera Manager.

Image Management

Images can be preloaded on all NodeManager hosts or they can be implicitly pulled at runtime if they are available in a public Docker registry, such as Docker hub. If the image does not exist on the NodeManager host and cannot be pulled, the container fails.

Docker Bind Mounted Volumes

Files and directories from the host are commonly needed within the Docker containers, which Docker provides through volumes. Examples include localized resources, Apache Hadoop binaries, and sockets. In order to make use of this feature, the following must be configured:

The administrator must define the volume whitelist by setting docker.allowed.ro-mounts and docker.allowed.rw-mounts to the list of parent directories that are allowed to be mounted.

The application submitter requests the required volumes at application submission time using the YARN_CONTAINER_RUNTIME_DOCKER_MOUNTS environment variable.

The administrator supplied whitelist is defined as a comma separated list of directories that are allowed to be mounted into containers. The source directory supplied by the user must either match or be a child of the specified directory.

The user supplied mount list is defined as a comma separated list in the form source:destination:mode. This format is ensured when you use Cloudera Manager, as a validation error is displayed, if a mount does not satisfy this format. The source is the file or directory on the host. The destination is the path within the container where the source will be bind mounted. The mode defines the mode the user expects for the mount, which can be “ro” or “rw” representing read-only and read-write mode respectively.