[YARN-9083] Support remote directory localization in yarn native service

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Assignee: Zhankun Tang
• Reporter: Zhankun Tang
• Created: 05/Dec/18 02:51
• Updated: 05/Dec/18 02:53

Description

When refining YARN-8714, found that the YARN localizer seems can handle remote directory directly. In FSDownload.java#downloadAndUnpack, it uses "FileUtil.copy" which can handle directory. This ability is added by YARN-2185.

For testing purpose, I changed distributedShell's client to let it localize an HDFS directory "mydir" directly. 

Path p = new Path("hdfs:///user/yarn/submarine/jobs/tf-job-001/staging" +
 "/mydir");
FileStatus scFileStatus = fs.getFileStatus(p);
LocalResource r =
 LocalResource.newInstance(URL.fromURI(p.toUri()),
 LocalResourceType.FILE, LocalResourceVisibility.APPLICATION,
 scFileStatus.getLen(), scFileStatus.getModificationTime());
localResources.put("mydir", r);

And YARN localizer indeed downloads the HDFS dir to local for DistributedShell.

yarn@master0-VirtualBox:~/apache-hadoop-install-dir/hadoop-dev-workspace$ ls /tmp/hadoop-yarn/nm-local-dir/usercache/yarn/appcache/application_1543898305524_0004/filecache/13/mydir
1.py  2.py  dir1  test_kill9.sh
yarn@master0-VirtualBox:~/apache-hadoop-install-dir/hadoop-dev-workspace$ ls /tmp/hadoop-yarn/nm-local-dir/usercache/yarn/appcache/application_1543898305524_0004/container_1543898305524_0004_01_000001/ -l
total 20
lrwxrwxrwx 1 yarn hadoop  111 12月  5 10:08 AppMaster.jar -> /tmp/hadoop-yarn/nm-local-dir/usercache/yarn/appcache/application_1543898305524_0004/filecache/10/AppMaster.jar
lrwxrwxrwx 1 yarn hadoop  103 12月  5 10:08 mydir -> /tmp/hadoop-yarn/nm-local-dir/usercache/yarn/appcache/application_1543898305524_0004/filecache/13/mydir

But the YARN native service seems doesn't know this YARN localizer ability and blocked it.

2018-12-05 10:06:40,286 ERROR client.ApiServiceClient: srcFile=hdfs://192.168.50.191:9000/user/yarn/submarine/jobs/tf-job-001/staging/mydir is a directory, which is not supported.

We should enable this ability in yarn native service.