Details
-
Improvement
-
Status: Resolved
-
Minor
-
Resolution: Fixed
-
3.2.0
-
None
Description
YARN currently supports three different Linux container runtimes (default, docker, and javasandbox). However, it would be relatively straightforward to support arbitrary runtime implementations. This would enable easier experimentation with new and emerging runtime technologies (runc, containerd, etc.) without requiring a rebuild and redeployment of Hadoop.
This could be accomplished via a simple configuration change:
<property> <name>yarn.nodemanager.runtime.linux.allowed-runtimes</name> <value>default,docker,experimental</value> </property> <property> <name>yarn.nodemanager.runtime.linux.experimental.class</name> <value>com.somecompany.yarn.runtime.ExperimentalLinuxContainerRuntime</value> </property>
In this example, yarn.nodemanager.runtime.linux.allowed-runtimes would now allow arbitrary values. Additionally, yarn.nodemanager.runtime.linux.{RUNTIME_KEY}.class would indicate the LinuxContainerRuntime implementation to instantiate. A no-argument constructor should be sufficient, as LinuxContainerRuntime already provides an initialize() method.
DockerLinuxContainerRuntime.isDockerContainerRequested(Map<String, String> env) and JavaSandboxLinuxContainerRuntime.isSandboxContainerRequested() could be generalized to isRuntimeRequested(Map<String, String> env) and added to the LinuxContainerRuntime interface. This would allow DelegatingLinuxContainerRuntime to select an appropriate runtime based on whether that runtime claimed ownership of the current container execution.
For backwards compatibility, the existing values (default,docker,javasandbox) would continue to be supported as-is. Under the current logic, the evaluation order is javasandbox, docker, default (with default being chosen if no other candidates are available). Under the new evaluation logic, pluggable runtimes would be evaluated after docker and before default, in the order in which they are defined in the allowed-runtimes list. This will change no behavior on current clusters (as there would be no pluggable runtimes defined), and preserves behavior with respect to ordering of existing runtimes.