[
https://issues.apache.org/jira/browse/YARN-6357?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15937110#comment-15937110
]
Haibo Chen commented on YARN-6357:
----------------------------------
Thanks for the pointer [~varun_impala_149e]. Reading through the discussion
there, it now makes sense to me why a TimelineWriter.writesync() is never added
(https://issues.apache.org/jira/browse/YARN-3949?focusedCommentId=14640959&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14640959).
However, I think there is still some confusion in the current TimelienWriter
API. Given the way TimelineCollector uses TimelineWriter, i.e. call flush()
right after write() for synchronous putEntities requests, TimelineCollector
expects TimelineWriter.write() to be asynchronous. In the meantime, it also
expects a TimelineResponse from this asynchronous method, which I think is
awkward and can cause confusion for alternative TimelineWriter implementations.
I propose, we replace TimelineResponse with void as the return type,
effectively making TimelineWriter API very much similar to that of
BufferedMutator. This way, for truly asynchronous implementations of
TimelineWriter.write(), such as HBaseTimelineWriter, no bogus TimelineResponse
is created any more. For synchronous implementations, response can be given
back in terms of IOException, that is, if the underlying synchronous store
succeeds, no exception is thrown, otherwise, the failure is wrapped in a
IOException and given back to TimelineCollector.
> Implement TimelineCollector#putEntitiesAsync
> --------------------------------------------
>
> Key: YARN-6357
> URL: https://issues.apache.org/jira/browse/YARN-6357
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: ATSv2, timelineserver
> Affects Versions: YARN-2928
> Reporter: Joep Rottinghuis
> Assignee: Haibo Chen
> Labels: yarn-5355-merge-blocker
> Attachments: YARN-6357.01.patch, YARN-6357.02.patch
>
>
> As discovered and discussed in YARN-5269 the
> TimelineCollector#putEntitiesAsync method is currently not implemented and
> TimelineCollector#putEntities is asynchronous.
> TimelineV2ClientImpl#putEntities vs TimelineV2ClientImpl#putEntitiesAsync
> correctly call TimelineEntityDispatcher#dispatchEntities(boolean sync,...
> with the correct argument. This argument does seem to make it into the
> params, and on the server side TimelineCollectorWebService#putEntities
> correctly pulls the async parameter from the rest call. See line 156:
> {code}
> boolean isAsync = async != null && async.trim().equalsIgnoreCase("true");
> {code}
> However, this is where the problem starts. It simply calls
> TimelineCollector#putEntities and ignores the value of isAsync. It should
> instead have called TimelineCollector#putEntitiesAsync, which is currently
> not implemented.
> putEntities should call putEntitiesAsync and then after that call
> writer.flush()
> The fact that we flush on close and we flush periodically should be more of a
> concern of avoiding data loss; close in case sync is never called and the
> periodic flush to guard against having data from slow writers get buffered
> for a long time and expose us to risk of loss in case the collector crashes
> with data in its buffers. Size-based flush is a different concern to avoid
> blowing up memory footprint.
> The spooling behavior is also somewhat separate.
> We have two separate methods on our API putEntities and putEntitiesAsync and
> they should have different behavior beyond waiting for the request to be sent.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
