Steve Loughran commented on SPARK-5925:

looking at this, I'm confused about what I'd written myself.

I concur with Marco, with one little point: if this is the YARN UI, then 
somewhere under YARN-896 there was discussion on moving the view on from 
"percentage done" —valid for simple batch runs— to, for long lived services, 
"percentage in use". That is, if the Spark AM knows its desired cluster size, 
it can tell the YARN RM that it's at 100% capacity; if below that it could set 
a different value. That is, the field will move from "percentage done" to 
"percentage that cluster size matches desired".

There's no change in the YARN API needed here; the main changes are (a) YARN RM 
to display allocation value of long-lived services different from short-lived 
executions (different text, colour), and for the Spark AM to update the RM with 
the current percentage value, changing it as containers come and go. 

> YARN - Spark progress bar stucks at 10% but after finishing shows 100%
> ----------------------------------------------------------------------
>                 Key: SPARK-5925
>                 URL: https://issues.apache.org/jira/browse/SPARK-5925
>             Project: Spark
>          Issue Type: Improvement
>          Components: YARN
>    Affects Versions: 1.2.1
>            Reporter: Laszlo Fesus
>            Priority: Minor
> I did set up a yarn cluster (CDH5) and spark (1.2.1), and also started Spark 
> History Server. Now I am able to click on more details on yarn's web 
> interface and get redirected to the appropriate spark logs during both job 
> execution and also after the job has finished. 
> My only concern is that while a spark job is being executed (either 
> yarn-client or yarn-cluster), the progress bar stucks at 10% and doesn't 
> increase as for MapReduce jobs. After finishing, it shows 100% properly, but 
> we are loosing the real-time tracking capability of the status bar. 
> Also tested yarn restful web interface, and it retrieves again 10% during 
> (yarn) spark job execution, and works well again after finishing. (I suppose 
> for the while being I should have a look on Spark Job Server and see if it's 
> possible to track the job via its restful web interface.)
> Did anyone else experience this behaviour? Thanks in advance.

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org

Reply via email to