[
https://issues.apache.org/jira/browse/HIVE-16104?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15922804#comment-15922804
]
Siddharth Seth commented on HIVE-16104:
---------------------------------------
Mostly looks good.
{code}
// TODO: not clear if we should wait for our last victim, or this. CR feedback.
{code}
Didn't get this comment. Should we wait for the specific preemption? - no, if
some other fragment completes.
If another fragment completes, we'll forget about what was preempted, and when
it was preempted. Don't think we should try fixing this corner case here.
FixedClock - May want to use "ControlledClock" which is already available.
The TODO on the thread should go.
On the test, think we should increase the sleep time + timeouts, or move to a
mechanism where the sleep is not required (that's where the clock was supposed
to help). 500ms typically causes a problems on slow/busy boxes, and will result
in test flakiness. A direct notifyAll to break out of the condition is ideal -
that gets into the area of whether the main code should be modified for tests.
Will probably require a wrapper wait interface which the test can control...
IAC, increased timeouts is probably the simplest at the moment.
> LLAP: preemption may be too aggressive if the pre-empted task doesn't die
> immediately
> -------------------------------------------------------------------------------------
>
> Key: HIVE-16104
> URL: https://issues.apache.org/jira/browse/HIVE-16104
> Project: Hive
> Issue Type: Bug
> Reporter: Sergey Shelukhin
> Assignee: Sergey Shelukhin
> Attachments: HIVE-16104.01.patch, HIVE-16104.02.patch,
> HIVE-16104.03.patch, HIVE-16104.04.patch, HIVE-16104.patch
>
>
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
