[
https://issues.apache.org/jira/browse/SPARK-14597?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15241358#comment-15241358
]
Mario Briggs commented on SPARK-14597:
--------------------------------------
I think there is an opportunity to merge both your approaches above. Let me
explain, taking how the onOutputOperationStarted/onOutputOperationCompleted is
already implemented.
So rather than providing a single time metric and a single start/complete event
that encompasses the generateJob for all OutputStreams, you could provide a
start/complete event for each individual outputstream generateJob and
onBatchComplete provide the metric for generateJob of all OutputStreams. This
way a user can also figure out if a individual outputstream is the culprit.
The above would require 2 additional things - pass an eventLoop to
DStreamGraph.generateJobs() method. This eventLoop should not be the existing
eventLoop instance in JobGenerator, but rather another new eventLoop instance
(say genJobEventLoop) in JobGenerator. This is because the existing
JobGenerator.eventLoop instance's thread is used to actually drive the Job
Generation and making that thread do additional tasks will increase latency in
Streaming. This new 'genJobEventLoop' will handle a GenJobStarted and
GenJobCompleted event and use those events to fire corresponding events to the
ListenerBus and gather the generateJob metric for all outputStreams and set it
in the JobSet
> Streaming Listener timing metrics should include time spent in JobGenerator's
> graph.generateJobs
> ------------------------------------------------------------------------------------------------
>
> Key: SPARK-14597
> URL: https://issues.apache.org/jira/browse/SPARK-14597
> Project: Spark
> Issue Type: Improvement
> Components: Spark Core, Streaming
> Affects Versions: 1.6.1, 2.0.0
> Reporter: Sachin Aggarwal
> Priority: Minor
>
> While looking to tune our streaming application, the piece of info we were
> looking for was actual processing time per batch. The
> StreamingListener.onBatchCompleted event provides a BatchInfo object that
> provided this information. It provides the following data
> - processingDelay
> - schedulingDelay
> - totalDelay
> - Submission Time
> The above are essentially calculated from the streaming JobScheduler
> clocking the processingStartTime and processingEndTime for each JobSet.
> Another metric available is submissionTime which is when a Jobset was put on
> the Streaming Scheduler's Queue.
>
> So we took processing delay as our actual processing time per batch. However
> to maintain a stable streaming application, we found that the our batch
> interval had to be a little less than DOUBLE of the processingDelay metric
> reported. (We are using a DirectKafkaInputStream). On digging further, we
> found that processingDelay is only clocking time spent in the ForEachRDD
> closure of the Streaming application and that JobGenerator's
> graph.generateJobs
> (https://github.com/apache/spark/blob/branch-1.6/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobGenerator.scala#L248)
> method takes a significant more amount of time.
> Thus a true reflection of processing time is
> a - Time spent in JobGenerator's Job Queue (JobGeneratorQueueDelay)
> b - Time spent in JobGenerator's graph.generateJobs (JobSetCreationDelay)
> c - Time spent in JobScheduler Queue for a Jobset (existing schedulingDelay
> metric)
> d - Time spent in Jobset's job run (existing processingDelay metric)
>
> Additionally a JobGeneratorQueue delay (#a) could be due to either
> graph.generateJobs taking longer than batchInterval or other JobGenerator
> events like checkpointing adding up time. Thus it would be beneficial to
> report time taken by the checkpointing Job as well
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
