[ 
https://issues.apache.org/jira/browse/TEZ-3816?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16122452#comment-16122452
 ] 

Siddharth Seth edited comment on TEZ-3816 at 8/10/17 10:29 PM:
---------------------------------------------------------------

I believe a max of 2 attempts (for a single task) running is handled via the 
regular flow? So that this does not end up spawning new attempts up to the max 
limit each time this timeout is exceeded.

May be easier to put all the logic in a single place - 
LegacySpeculator.maybeScheduleASpeculation, instead of having to change 
multiple methods on the LegactyRuntimeEstimator class.

Another possible enhancement, at a later point, would be to speculate earlier 
rather than later, once there is sufficient confidence that the current attempt 
will miss the timeout, based on progress information available.

ALso, a test would be useful.


was (Author: sseth):
I believe a max of 2 attempts (for a single task) running is handled via the 
regular flow? So that this does not end up spawning new attempts up to the max 
limit each time this timeout is exceeded.

May be easier to put all the logic in a single place - 
LegacySpeculator.maybeScheduleASpeculation, instead of having to change 
multiple methods on the LegactyRuntimeEstimator class.

Another possible enhancement, at a later point, would be to speculate earlier 
rather than later, once there is sufficient confidence that the current attempt 
will miss the timeout, based on progress information available.

> Ability to automatically speculate single-task vertices
> -------------------------------------------------------
>
>                 Key: TEZ-3816
>                 URL: https://issues.apache.org/jira/browse/TEZ-3816
>             Project: Apache Tez
>          Issue Type: Improvement
>            Reporter: Muhammad Samir Khan
>            Assignee: Muhammad Samir Khan
>         Attachments: TEZ-3816.001.patch
>
>
> When a single-task vertex is unlucky, it lands on a very slow node. 
> Speculation doesn't currently apply when there are no other tasks to compare 
> with. It would be good to have a configurable timeout after which the tasks 
> automatically speculate.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Reply via email to