[jira] [Created] (YARN-4970) Difficult to trace "Connection Refused" in AM Proxying

["Matthew Byng-Maddick (JIRA)" <ji...@apache.org>]

Matthew Byng-Maddick created YARN-4970:
------------------------------------------

             Summary: Difficult to trace "Connection Refused" in AM Proxying
                 Key: YARN-4970
                 URL: https://issues.apache.org/jira/browse/YARN-4970
             Project: Hadoop YARN
          Issue Type: Improvement
          Components: webapp
    Affects Versions: 2.6.0
         Environment: Hadoop-2.6.0-CDH5.5.1, linux
            Reporter: Matthew Byng-Maddick
            Priority: Minor


In generating an HA YARN config (effectively using my own tools for generation), I missed out the multiple specification of {{yarn.resourcemanager.webapp.address}}, which produced a "Connection Refused" similar to that in YARN-800 (because of a misunderstanding about what the proxy address was) from the AM webapp.

It occurs to me, though, in tracing the code, that the behaviour of {{hadoop-yarn-common/src/main/java/org/apache/hadoop/yarn/util/RMHAUtils.java}}: {{getRMHAWebappAddresses(YarnConfiguration)}} should probably mirror the effect of the startup of the resourcemanager itself, such that if the id-suffixed key doesn't exist, you end up using the {{yarn.resourcemanager.address}} with the port number replaced by the one taken from {{yarn.resourcemanager.webapp.address}} (or its https equivalent). That is, if you had a config like:
{code}
<configuration>
  <property><name>yarn.resourcemanager.ha.enabled</name>
    <value>true</value></property>
  <property><name>yarn.resourcemanager.ha.rm-ids</name>
    <value>hosta,hostb</value></property>
  <property><name>yarn.resourcemanager.webapp.address</name>
    <value>0.0.0.0:8088</value></property>
  <property><name>yarn.resourcemanager.address.hosta</name>
    <value>hosta:8032</value></property>
  <property><name>yarn.resourcemanager.address.hostb</name>
    <value>hostb:8032</value></property>
</configuration>
{code}
you would end up with ("{{hosta:8088}}", "{{hostb:8088}}") as the {{List<String>}} result of the function above, rather than the current empty list result.

This would certainly give a better principle of least-surprise for operators of yarn (especially those like myself who ended up configuring it wrongly).

Thoughts? Reasons why this isn't a good idea? (I'm afraid my Java's a bit read-only which is why I haven't suggested a patch)



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)