[
https://issues.apache.org/jira/browse/SPARK-25645?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16639024#comment-16639024
]
Marcelo Vanzin commented on SPARK-25645:
----------------------------------------
This seems similar to SPARK-24787. Have you tried with hflush instead of hsync?
Some flushing might be nice, perhaps with an option to configure things (e.g.
none/job/stage/all for different amounts of flushing).
> Add provision to disable EventLoggingListener default flush/hsync/hflush for
> all events
> ---------------------------------------------------------------------------------------
>
> Key: SPARK-25645
> URL: https://issues.apache.org/jira/browse/SPARK-25645
> Project: Spark
> Issue Type: Improvement
> Components: Spark Core
> Affects Versions: 2.3.2
> Reporter: Devaraj K
> Priority: Major
>
> {code:java|title=EventLoggingListener.scala|borderStyle=solid}
> private def logEvent(event: SparkListenerEvent, flushLogger: Boolean = false)
> {
> val eventJson = JsonProtocol.sparkEventToJson(event)
> // scalastyle:off println
> writer.foreach(_.println(compact(render(eventJson))))
> // scalastyle:on println
> if (flushLogger) {
> writer.foreach(_.flush())
> hadoopDataStream.foreach(ds => ds.getWrappedStream match {
> case wrapped: DFSOutputStream =>
> wrapped.hsync(EnumSet.of(SyncFlag.UPDATE_LENGTH))
> case _ => ds.hflush()
> })
> }
> {code}
> There are events which come with flushLogger=true and go through the
> underlying stream flush, Here I tried running apps with disabling the
> flush/hsync/hflush for all events and see that there is significant
> improvement in the app completion time and also there are no event drops,
> posting more details in the comments section.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
