Github user sryza commented on the pull request:

    https://github.com/apache/spark/pull/2746#issuecomment-59956545
  
    > Even for longer-running tasks, we still have a problem. If the add timer 
expires multiple times while these tasks are being run, we may end up double 
counting the number of executors needed.
    
    My assumption was that we would keep the "target" number of executors equal 
to or less than the number of tasks within the currently running stages, which 
would solve this problem, IIUC.
    
    That said, I'll stop arguing for Kay's approach.  The exponential approach 
sounds very reasonable to me as well.
    
    What I do feel really strongly about is that we shouldn't add new 
configuration knobs that are required to get decent utilization for the average 
app.  Most of the users I've spoken to find a config as simple as 
`--executor-cores` difficult to reason about.  And even if they completely 
grasp how to set it, it's just one other thing they have to think about that 
distracts them from the job they're actually trying to complete.  This is one 
of the core complaints form Alex Rubinsteyn's infamous [blog 
post](http://blog.explainmydata.com/2014/05/spark-should-be-better-than-mapreduce.html).
  I think some simple rules could remove this burden from the user.  For 
example, we could cap the number of outstanding executor requests at the number 
that would be able to handle all the pending tasks - there's no reason it 
should ever need to exceed this.  A simple way to explain this to a user would 
be that Spark will never try to acquire more resources than it would need to 
run all the work it'
 s ready to run at this moment.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to