[
https://issues.apache.org/jira/browse/TEZ-1032?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14015979#comment-14015979
]
Rajesh Balamohan commented on TEZ-1032:
---------------------------------------
Corner Case:
========
1. User specifies v1[1], v2[10] and let us assume 200 tasks are there in V1.
2. Let us assume user has specified same profiler JVM option. (e.g
“-agentpath:/opt/yourkit/bin/linux-x86-64/libyjpagent.so”).
3. All profiled tasks would have the same JVM opts
4. Task 1 gets allocated to container-1. Non profiled tasks would not be
allocated to this container, as the JVM opts are different.
5. However, task 10 of v2 which has same profiler option can get allocated to
container-1.
- Even though this is a rare case, this can happen and it would end up
generating wrong profiler snapshot which has a mix of v1 and v2. This
definitely needs to be avoided, if we need to nail down certain bottlenecks.
Easy workaround for this corner case is that, user can disable container reuse
for the job being profiled. Thoughts?
> allow specifying tasks/vertices to be profiled
> ----------------------------------------------
>
> Key: TEZ-1032
> URL: https://issues.apache.org/jira/browse/TEZ-1032
> Project: Apache Tez
> Issue Type: Improvement
> Reporter: Sergey Shelukhin
> Assignee: Rajesh Balamohan
> Attachments: TEZ-1032-v1.patch, TEZ-1032-v4.patch, TEZ-1032-v5.patch,
> TEZ-1032-v6.patch, TEZ-1032-v8.patch, TEZ-1032-v9.patch
>
>
> Specifying container numbers (via tez.profile.container.list) is not very
> useful... it would be nice to be able to specify particular stage (e.g. "Map
> 6") to profile.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
