[ 
https://issues.apache.org/jira/browse/OOZIE-1118?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13611241#comment-13611241
 ] 

Virag Kothari commented on OOZIE-1118:
--------------------------------------

I was thinking of the case where the coordinators are configured to be purged 
in 3 months while the workflows are configured to be purged in lets say 6 
months. Then once the coordinator becomes eligible to purge, the purge service 
will bring all the workflows of that coordinator in memory every purge interval 
till 3 months and fail to purge. However, as it a system configuration probably 
its fine and all this validation may be added later.

The patch looks fine to me.
One question on the patch:
For checking eligibility of purge, all the children are brought into memory. 
Instead of that, is it better to filter at the db side?
E.g Having an sql like 
{code}
select count(*) from child c where c.parent = parent and c is not eligible to 
purge.
if (count > 0) {
 isEligibleToPurge = false
}
else {
  Purge the parent and children
}
{code}
Also, all the queries should bring only selected columns instead of select *. 
Oozie had run into OOM issues due to select * before.


                
> improve logic of purge service
> ------------------------------
>
>                 Key: OOZIE-1118
>                 URL: https://issues.apache.org/jira/browse/OOZIE-1118
>             Project: Oozie
>          Issue Type: Improvement
>          Components: bundle, coordinator, workflow
>    Affects Versions: 3.3.0
>            Reporter: Alejandro Abdelnur
>            Assignee: Robert Kanter
>             Fix For: trunk
>
>         Attachments: OOZIE-1118.patch, OOZIE-1118.patch, OOZIE-1118.patch, 
> OOZIE-1118.patch, OOZIE-1118.patch, OOZIE-1118.patch, OOZIE-1118.patch
>
>
> The current logic of the purge service is flat. I.e., WF purging only takes 
> into account WF end time, it does not take into account that the WF was 
> started by a COORD job. This means that completed WFs of a running COORD job 
> could be purge if the COORD job runs for longer that the purge age.
> One way of addressing this would be:
> * WF purging only purges WF jobs started directly by a client call.
> * COORD purging purges COORD jobs started directly by a client call. It also 
> purges the WF jobs created by the COORD jobs being purged.
> * BUNDLE purging purges BUNDLE jobs, and the corresponding COORD jobs and WF 
> jobs.
> This could be handled by a new property in the job beans 'job-owner'. Set to 
> 'self' it would mean it can be purged by the same job type purger. If set to 
> other value, then it is a higher level purger the one responsible for purging 
> it.
> This means that for a WF job started by COORD job started by a BUNDLE job, 
> the WF job and the COORD job would have the BUNDLE job as owner, while the 
> BUNDLE with have 'self' as owner.
> This ownership propagation would also have
> A caveat here would be how to handle sub-workflows. 
> I guess we should check if the wf was created from  coord, and if then let 
> the coord purge take care of that, meaning wf purge does not purge wf started 
> by coords.
> Similarly, the same should also apply for sub-WFs.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Reply via email to