[YARN-11386] Fix issue with classpath resolution

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 3.4.0
• Fix Version/s: 3.4.0
• Component/s: nodemanager
• Labels: pull-request-available

Environment

Windows 10

• Target Version/s: 3.4.0
• Hadoop Flags: Reviewed

Language

Java

• Assignee: Gautham Banasandra
• Reporter: Gautham Banasandra
• Due: 11/Dec/22
• Created: 04/Dec/22 10:25
• Updated: 28/Jan/24 02:05
• Resolved: 07/Dec/22 14:59

Description

While launching the AM, the client puts the classpath into environment HashMap of the ContainerLaunchContext. To support cross-platform compatibility, some special notations must be used by the client. For example, <CPS> must be used as the classpath separator instead of ";" (Windows) or ":" (Linux). The NM would resolve all the <CPS> using the appropriate path separator according to the OS platform, prior to launching the container. In addition to this, the NM is also expected to expand and resolve all of the wildcard ( * ) occurrences in the classpath before launching the container, since the JVM doesn't understand the wildcard charater.

The issue we see here is, neither <CPS> resolution nor wildcard expansion is happening.

I tested this from Spark. To run a Spark application on YARN, one needs to set the spark.yarn.jars config to point to the location where Spark jars are present. NM would then add these jars to the classpath before launching the Spark AM container. The resolution didn't happen -

Manifest-Version:
1.0
Class-Path:
file:/D:/tmp/hadoop-Gautham/nm-local-dir/usercache/Gautham/appcache/application_1670138957874_0001/container_1670138957874_0001_01_000001/%7B%7BPWD%7D%7D%3CCPS%3E%7B%7BPWD%7D%7D/__spark_conf__%3CCPS%3E%7B%7BPWD%7D%7D/__spark_libs__/*%3CCPS%3E/D:/projects/github/apache/spark/jars/*%3CCPS%3E%7B%7BPWD%7D%7D/__spark_conf__/__hadoop_conf__
file:/D:/tmp/hadoop-Gautham/nm-local-dir/usercache/Gautham/appcache/application_1670138957874_0001/container_1670138957874_0001_01_000001/__spark_conf__/
file:/D:/tmp/hadoop-Gautham/nm-local-dir/usercache/Gautham/appcache/application_1670138957874_0001/container_1670138957874_0001_01_000001/__app__.jar

Note that %3CCPS%3E in the above code block should've been replaced by ";" or ":" as part of classpath resolution. Since this didn't happen, NM failed to launch the Spark AM container -

Could not find or load main class org.apache.spark.deploy.yarn.ApplicationMaster

Issue Links

• is caused by: YARN-10664 Allow parameter expansion in NM_ADMIN_USER_ENV
• links to: GitHub Pull Request #5183