Sangjin Lee commented on YARN-3166:

[~gtCarrera9], sorry for my delayed response.

It looks good mostly. I have some feedback:
- As [~zjshen] mentioned, we need to sort out the RM/NM dependency on the 
timelineservice module. The NM dependency is more of a fluke, but we need to 
think about the RM dependency because it needs to start its own aggregator 
service. I believe [~Naganarasimha] mentioned this in another JIRA. Perhaps 
this is unavoidable if RM is going to start the aggregator? I am not aware of 
any clean pluggable service mechanism for RM (like the aux services for NM). 
Another idea if we don't want that is to move the base aggregator class into 
- I think as a rule, it would be good to make sure not to disturb the old ATS 
classes. IIUC we're deprecating the old ATS classes, but we're not going to 
modify them in an incompatible way (e.g. moving classes, removing classes, 
changing interfaces, etc.), as that would be extremely disruptive once this is 
- What is the difference between TimelineStorage and TimelineStorageImpl?

> [Source organization] Decide detailed package structures for timeline service 
> v2 components
> -------------------------------------------------------------------------------------------
>                 Key: YARN-3166
>                 URL: https://issues.apache.org/jira/browse/YARN-3166
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>            Reporter: Li Lu
>            Assignee: Li Lu
> Open this JIRA to track all discussions on detailed package structures for 
> timeline services v2. This JIRA is for discussion only.
> For our current timeline service v2 design, aggregator (previously called 
> "writer") implementation is in hadoop-yarn-server's:
> {{org.apache.hadoop.yarn.server.timelineservice.aggregator}}
> In YARN-2928's design, the next gen ATS reader is also a server. Maybe we 
> want to put reader related implementations into hadoop-yarn-server's:
> {{org.apache.hadoop.yarn.server.timelineservice.reader}}
> Both readers and aggregators will expose features that may be used by YARN 
> and other 3rd party components, such as aggregator/reader APIs. For those 
> features, maybe we would like to expose their interfaces to 
> hadoop-yarn-common's {{org.apache.hadoop.yarn.timelineservice}}? 
> Let's use this JIRA as a centralized place to track all related discussions. 

This message was sent by Atlassian JIRA

Reply via email to