[
https://issues.apache.org/jira/browse/SPARK-14597?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15250661#comment-15250661
]
Sachin Aggarwal commented on SPARK-14597:
-----------------------------------------
hi Mario,
I have extended the approach2 added a new parameter to job class to capture the
job creation time delay, once the job gets created, I set time taken to create
each job and user can get this information in StreamingListener methods
onOutputOperationStarted and onOutputOperationCompleted corresponding to each
job,
for batch level data user can use
batchCompleted.batchInfo.batchJobSetCreationDelay in onBatchCompleted method of
StreamingListener
> 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]
