[YARN-9014] runC container runtime

• Type: New Feature
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Labels: Docker

Release Note

The new RuncContainerRuntime is an experimental feature that allows Hadoop tasks to be run inside of runC containers. This feature is similar to the DockerLinuxContainerRuntime, but is catered more to Hadoop. Notably, the RuncContainerRuntime leverages squashFS for its image format and spreads the images across all disks accessible to Hadoop using the distributed cache. Additionally, unlike Docker, the RuncContainerRuntime does not rely on any daemons.

• Assignee: Eric Badger
• Reporter: Jason Darrell Lowe
• Created: 12/Nov/18 22:28
• Updated: 09/Jan/20 16:53

Description

This JIRA tracks a YARN container runtime that supports running containers in images built by Docker but the runtime does not use Docker directly, and Docker does not have to be installed on the nodes. The runtime leverages the OCI runtime standard to launch containers, so an OCI-compliant runtime like runc is required. runc has the benefit of not requiring a daemon like dockerd to be running in order to launch/control containers.

The layers comprising the Docker image are uploaded to HDFS as squashfs images, enabling the runtime to efficiently download and execute directly on the compressed layers. This saves image unpack time and space on the local disk. The image layers, like other entries in the YARN distributed cache, can be spread across the YARN local disks, increasing the available space for storing container images on each node.

A design document will be posted shortly.

Sub-Tasks

• Restructure DockerLinuxContainerRuntime to extend a new OCIContainerRuntime: Resolved
• Add C changes for the new RuncContainerRuntime: Resolved
• Add Java changes for the new RuncContainerRuntime: Resolved
• Create docker-to-squash tool for image conversion: Patch Available
• Port YARN-8569 to RuncContainerRuntime: Open
• Make container-executor mount logic modular: Resolved
• Investigate image/tag management and how to do docker-to-squash conversion at scale: Open
• Work around hard-coded tmp and /var/tmp bind-mounts in the container's working directory: Open
• Add relaunchContainer logic to RuncContainerRuntime: Open
• Add execContainer logic to RuncContainerRuntime: Open
• Add networking support other than host to RuncContainerRuntime: Open
• Add getIpAndHost to RuncContainerRuntime: Open
• RunC container repository v2: Open
• CLI tool for docker-to-squashfs conversion (pure Java): Open