[
https://issues.apache.org/jira/browse/YARN-8725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16620834#comment-16620834
]
Zhankun Tang commented on YARN-8725:
------------------------------------
[~leftnoteasy]
{quote}cleanup whole staging dir seems overkill because models, etc. by default
is placed under the directory as well.
{quote}
Thanks for pointing out this model directory stuff. This patch seems overkill.
Let's hold on it.
I double-checked the relations between "--checkpoint_path",
"--saved_model_path" and "–input_path".
For the --checkpoint_path", the situations are as below:
1. if set, in my case I set this "--checkpoint_path
hdfs://default/user/yarn/cifar-10-jobdir". So it'll be safe to delete the
staging dir since the checkout_dir that contains model data is not in staging
dir.
2. if not set, it'll replace "%checkout_path" with
"submarine/jobs/tf-job-001/staging/checkpoint_path" due to below code:
{code:java}
public static String replacePatternsInLaunchCommand(String specifiedCli,
RunJobParameters jobRunParameters,
RemoteDirectoryManager directoryManager) throws IOException {
String jobDir = jobRunParameters.getCheckpointPath();
if (null == jobDir) {
jobDir = directoryManager.getJobCheckpointDir(jobRunParameters.getName(),
true).toString();
}
String input = jobRunParameters.getInputPath();
String savedModelDir = jobRunParameters.getSavedModelPath();
if (null == savedModelDir) {
savedModelDir = jobDir;
}
Map<String, String> replacePattern = new HashMap<>();
if (jobDir != null) {
replacePattern.put("%" + CliConstants.CHECKPOINT_PATH + "%", jobDir);
}
...
if (savedModelDir != null) {
replacePattern.put("%" + CliConstants.SAVED_MODEL_PATH + "%",
savedModelDir);
}{code}
{code:java}
2018-09-19 23:19:34,729 INFO yarnservice.YarnServiceJobSubmitter: Worker
command =[cd /cifar10_estimator && python cifar10_main.py
--data-dir=hdfs://default/user/yarn/cifar-10-data
--job-dir=submarine/jobs/tf-job-001/staging/checkpoint_path --num-gpus=0
--train-steps=2]{code}
But the job failed due to invalid path passed to "--job-dir" per my testing. It
should be a URI start with "hdfs://".
And attached the script I use for this testing. Has submitted a Jira to track
this. YARN-8799
{code}
yarn jar
$HADOOP_BASE_DIR/home/share/hadoop/yarn/hadoop-yarn-submarine-3.2.0-SNAPSHOT.jar
job run \
-verbose \
-wait_job_finish \
-keep_staging_dir \
--env DOCKER_JAVA_HOME=/usr/lib/jvm/java-8-oracle \
--env DOCKER_HADOOP_HDFS_HOME=/hadoop-3.2.0-SNAPSHOT \
--name tf-job-001 \
--docker_image tangzhankun/tensorflow \
--input_path hdfs://default/user/yarn/cifar-10-data \
--worker_resources memory=4G,vcores=2 \
--worker_launch_cmd "cd /cifar10_estimator && python cifar10_main.py
--data-dir=%input_path% --job-dir=%checkpoint_path% --num-gpus=0
--train-steps=2"{code}
*I thought that if a user sets the "–checkoutpoint_path" option explicitly, it
seems not possible that the end user set a path string under the internal
staging dir like
"*hdfs*://default/user/yarn/submarine/jobs/*tf*-job-001/staging/checkpoint_path"(Because
the user is better to not know so such details). And in the other hand, don't
set this option but still use the pattern "%checkpoint_path" in worker command
seems strange to me.*
*So I assume we'll put checkpoint_dir outside of staging dir, missed above
script test and also the current code fact that we intend to put it under the
staging dir by default. :)*
For the "–input_path", it's a must option. Nothing more to discuss except we
should check invalid value(YARN-8798).
For the "--saved_model_path", it might have the same default value issue(needs
more tests). But it's mainly for serving. won't discuss here.
{quote}And logics in your patch cleans up dirs after job submitted. It is
possible that workers get launched after dir got deleted.
{quote}
Could you please elaborate on this?
{quote}I'm not sure if we can do many meaningful things here in the client
code. It might be better to do this in the server side, I don't have a clear
idea about how to handle the service part. Maybe it should be a plugin of
ApiServer, or it is a completely new service like a system service.
{quote}
Maybe let's talk about this offline.
> Submarine job staging directory has a lot of useless
> PRIMARY_WORKER-launch-script-***.sh scripts when submitting a job multiple
> times
> --------------------------------------------------------------------------------------------------------------------------------------
>
> Key: YARN-8725
> URL: https://issues.apache.org/jira/browse/YARN-8725
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Zac Zhou
> Assignee: Zhankun Tang
> Priority: Major
> Attachments: YARN-8725-trunk.001.patch
>
>
> Submarine jobs upload core-site.xml, hdfs-site.xml, job.info and
> PRIMARY_WORKER-launch-script****.sh to staging dir.
> The core-site.xml, hdfs-site.xml and job.info would be overwritten if a job
> is submitted multiple times.
> But PRIMARY_WORKER-launch-script****.sh would not be overwritten, as it has
> random numbers in its name.
> The files in the staging dir are as follows：
> {code:java}
> -rw-r----- 2 hadoop hdfs 580 2018-08-17 10:11
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/PRIMARY_WORKER-launch-script6954941665090337726.sh
> -rw-r----- 2 hadoop hdfs 580 2018-08-17 10:02
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/PRIMARY_WORKER-launch-script7037369696166769734.sh
> -rw-r----- 2 hadoop hdfs 580 2018-08-17 10:06
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/PRIMARY_WORKER-launch-script8047707294763488040.sh
> -rw-r----- 2 hadoop hdfs 15225 2018-08-17 18:46
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/PRIMARY_WORKER-launch-script8122565781159446375.sh
> -rw-r----- 2 hadoop hdfs 580 2018-08-16 20:48
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/PRIMARY_WORKER-launch-script8598604480700049845.sh
> -rw-r----- 2 hadoop hdfs 580 2018-08-17 14:53
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/PRIMARY_WORKER-launch-script971703616848859353.sh
> -rw-r----- 2 hadoop hdfs 580 2018-08-17 10:16
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/PRIMARY_WORKER-launch-script990214235580089093.sh
> -rw-r----- 2 hadoop hdfs 8815 2018-08-27 15:54
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/core-site.xml
> -rw-r----- 2 hadoop hdfs 11583 2018-08-27 15:54
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/hdfs-site.xml
> -rw-rw-rw- 2 hadoop hdfs 846 2018-08-22 10:56
> hdfs://submarine/user/hadoop/submarine/jobs/standlone-tf/staging/job.info
> {code}
>
> We should stop the staging dir from growing or have a way to clean it up
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
