[ https://issues.apache.org/jira/browse/BEAM-1048?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15693377#comment-15693377
]
Amit Sela commented on BEAM-1048:
---------------------------------
To clarify - calling `rdd.count()` on the read input RDD creates a batch job each interval.
I was following "direct kafka" implementation, but in that case it's only metadata computation
as the read from Kafka is known ahead of actual read (it has a preset [start, end] offsets).
This is used for UI visibility, and feeding the {{RateController}}.
If {{RateController}} is not used, we might want to disable this for now.. (UI will lose count
visualisation too).
[~ksalant] I've assigned you since you're working on it. If resolving this prolongs, let's
open a sub-task to disable this if {{RateController}} is not used, and enable once resolved.
> Spark Runner streaming batch duration does not include duration of reading from source
> ---------------------------------------------------------------------------------------
>
> Key: BEAM-1048
> URL: https://issues.apache.org/jira/browse/BEAM-1048
> Project: Beam
> Issue Type: Bug
> Components: runner-spark
> Affects Versions: 0.4.0-incubating
> Reporter: Kobi Salant
> Assignee: Kobi Salant
>
> Spark Runner streaming batch duration does not include duration of reading from source
this is because we perform rdd.count in SparkUnboundedSourcewhich that invokes a regular spark
job outside the streaming context.
> We do it for reporting the batch size both for UI and back pressure
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
|