[ https://issues.apache.org/jira/browse/YARN-3372?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14369361#comment-14369361 ]
Steve Loughran commented on YARN-3372: -------------------------------------- what naming scheme do you propose? SLIDER uses that of (yarn-app-class, yarn-app-name) with the implicit username, and rely on a policy of "one unique name". I guess that gets confused over time, even if enforced. Something like a uuid would be good -though it needs to valid DNS name for future DNS support., such as {{uuid0005455feaf}}. For YARN-launched apps, the yarn app ID is guaranteed to be unique for the lifespan of the RM/cluster; attempt ID across attempts > Collision-free unique bindings & refresh APIs for service records > ----------------------------------------------------------------- > > Key: YARN-3372 > URL: https://issues.apache.org/jira/browse/YARN-3372 > Project: Hadoop YARN > Issue Type: Sub-task > Components: yarn > Reporter: Gopal V > > The current bind() operation binds to a hard entry name for the service > record, which makes it impossible for a truly distributed application without > a centralized service to register without pre-determined naming conventions. > The uniqueness does not need to guarantee ordering or any other leakage of > abstractions, merely that each bind() returns a unique path the record was > bound to. And that the TTL refresh can periodically update that exact record > as an active API. > These are state-less auto-configuration mechanisms inspired by the IPv6 > improvements over DNS for resolution. Instead of relying ICMPv6, this uses > the registry to keep a collective memory of unique identities to which > endpoints are delegated to. > This is only obliquely related to the Slider registration as even those do > not track the generational ids for restarted daemons from the same > container-id. -- This message was sent by Atlassian JIRA (v6.3.4#6332)