[
https://issues.apache.org/jira/browse/YARN-7399?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Eric Yang updated YARN-7399:
----------------------------
Description:
In Slider, metadata is stored in user's home directory. Slider command line
interface interacts with HDFS directly to list deployed applications and invoke
YARN API or HDFS API to provide information to user. This design works for a
single user manage his/her own applications. When this design has been ported
to Yarn services, it becomes apparent that this design is difficult to list all
deployed applications on Hadoop cluster for administrator to manage
applications. Resource Manager needs to crawl through every user's home
directory to compile metadata about deployed applications. This can trigger
high load on namenode to list hundreds or thousands of list directory calls
owned by different users. Hence, it might be best to centralize the metadata
storage to Solr or HBase to reduce number of IO calls to namenode for manage
applications.
In Slider, one application is composed of metainfo, specifications in json, and
payload of zip file that contains application code and deployment code. Both
meta information, and zip file payload are stored in the same application
directory in HDFS. This works well for distributed applications without central
application manager that oversee all application.
In the next generation of application management, we like to centralize
metainfo and specifications in json to a centralized storage managed by YARN
user, and keep the payload zip file in user's home directory or in docker
registry. This arrangement can provide a faster lookup for metainfo when we
list all deployed applications and services on YARN dashboard.
When we centralize metainfo to YARN user, we also need to build ACL to enforce
who can manage applications, and make update. The current proposal is:
yarn.admin.acl - list of groups that can submit/reconfigure/pause/kill all
applications
normal users - submit/reconfigure/pause/kill his/her own applications
was:In Slider, metadata is stored in user's home directory. Slider command
line interface interacts with HDFS directly to list deployed applications and
invoke YARN API or HDFS API to provide information to user. This design works
for a single user manage his/her own applications. When this design has been
ported to Yarn services, it becomes apparent that this design is difficult to
list all deployed applications on Hadoop cluster for administrator to manage
applications. Resource Manager needs to crawl through every user's home
directory to compile metadata about deployed applications. This can trigger
high load on namenode to list hundreds or thousands of list directory calls
owned by different users. Hence, it might be best to centralize the metadata
storage to Solr or HBase to reduce number of IO calls to namenode for manage
applications.
> Yarn services metadata storage improvement
> ------------------------------------------
>
> Key: YARN-7399
> URL: https://issues.apache.org/jira/browse/YARN-7399
> Project: Hadoop YARN
> Issue Type: Improvement
> Components: yarn-native-services
> Reporter: Eric Yang
>
> In Slider, metadata is stored in user's home directory. Slider command line
> interface interacts with HDFS directly to list deployed applications and
> invoke YARN API or HDFS API to provide information to user. This design works
> for a single user manage his/her own applications. When this design has been
> ported to Yarn services, it becomes apparent that this design is difficult to
> list all deployed applications on Hadoop cluster for administrator to manage
> applications. Resource Manager needs to crawl through every user's home
> directory to compile metadata about deployed applications. This can trigger
> high load on namenode to list hundreds or thousands of list directory calls
> owned by different users. Hence, it might be best to centralize the metadata
> storage to Solr or HBase to reduce number of IO calls to namenode for manage
> applications.
> In Slider, one application is composed of metainfo, specifications in json,
> and payload of zip file that contains application code and deployment code.
> Both meta information, and zip file payload are stored in the same
> application directory in HDFS. This works well for distributed applications
> without central application manager that oversee all application.
> In the next generation of application management, we like to centralize
> metainfo and specifications in json to a centralized storage managed by YARN
> user, and keep the payload zip file in user's home directory or in docker
> registry. This arrangement can provide a faster lookup for metainfo when we
> list all deployed applications and services on YARN dashboard.
> When we centralize metainfo to YARN user, we also need to build ACL to
> enforce who can manage applications, and make update. The current proposal is:
> yarn.admin.acl - list of groups that can submit/reconfigure/pause/kill all
> applications
> normal users - submit/reconfigure/pause/kill his/her own applications
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
