Steve Loughran commented on YARN-3067:

Doug: I agree this wold be good.

Furthermore, /conf isn't guaranteed to be useful by clients, as it can include 
things like 0 ports & wildcarded hosts which are bonded at startup time.

For Slider we have written a REST API which [serves up 
 -of both the AM and of applications we deploy (e.g hbase-site.xml)

Some aspects of it we like
# depending on the extension of the URL: .xml. json or .properties, the content 
comes back in the appropriate format
# you can ask for a single specific property by appending it to the file, 
getting it back or a 404.
# It registers itself in the YARN registry, so you can find it. It is designed 
to be reusable

* While it serves up the dynamic values of deployed apps, I think it still 
doesn't include any dynamically resolved ports/hosts used to deploy the 
AM..that's something I'd like fixed.
* You may want to specify a filter for secrets like s3 credentials

For a YARN r/o API, I'd like to see the same feature set, and have something 
standardised enough that Slider (and other things, like HBase, HDFS, ...) could 
also implement if they chose to serve up their configs the same way —and we'd 
have a reference implementation for them to use.

If you want to take what we've done and use as the basis for a YARN/re-usable 
API, I'll help.

> Read-only REST API for System Configuration
> -------------------------------------------
>                 Key: YARN-3067
>                 URL: https://issues.apache.org/jira/browse/YARN-3067
>             Project: Hadoop YARN
>          Issue Type: Improvement
>          Components: api, client
>    Affects Versions: 2.6.0
>         Environment: REST API
>            Reporter: Doug Haigh
> While the name node exposes its configuration via <nn_host>:<nn_port>/conf 
> and resource manager exposes its configuration via <rm_host>:<rm_port>/conf, 
> neither provide a complete picture of the system's configuration for 
> applications that use the YARN REST API.
> This JIRA is to request a REST API to get all services' configuration 
> information whether it is stored in the resource manager or something 
> external like zookeeper. Essentially this would return information similar to 
> what <nn>/conf and <rm>/conf return, but be guaranteed to have all 
> information so that a REST API user would not require Hadoop configuration 
> files to know how services are setup.

This message was sent by Atlassian JIRA

Reply via email to