[
https://issues.apache.org/jira/browse/TEZ-3696?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15993419#comment-15993419
]
Bikas Saha commented on TEZ-3696:
---------------------------------
Thanks for the ping. Looking at the code again, I am not sure why I had the
check for succeeded attempts for sending the completed event. I renamed the
event type from succeeded to completed in the same patch and hence I may have
intended to stop differentiating them. But the sending code for the completed
event was under the succeeded check. That seems inconsistent.
If the above is correct, then perhaps the issue would happen even without
speculation and on every attempt failure. Because a failed attempt would not
decrease the running count and so its retry would not get scheduled, leading to
an off-by-N situation in the concurrency count in the dag scheduler. If this is
correct, then perhaps the fix is only to send the completed event all the time
and not make any other changes in the dag scheduler itself.
> Jobs can hang when both concurrency and speculation are enabled
> ---------------------------------------------------------------
>
> Key: TEZ-3696
> URL: https://issues.apache.org/jira/browse/TEZ-3696
> Project: Apache Tez
> Issue Type: Bug
> Reporter: Eric Badger
> Assignee: Eric Badger
> Attachments: TEZ-3696.001.patch, TEZ-3696.002.patch,
> TEZ-3696.003.patch
>
>
> We can reproduce the hung job by doing the following:
> 1. Run a sleep job with a concurrency of 1, speculation enabled, and 3 tasks
> {noformat}
> HADOOP_CLASSPATH="$TEZ_HOME/*:$TEZ_HOME/lib/*:$TEZ_CONF_DIR" yarn jar
> $TEZ_HOME/tez-tests-*.jar mrrsleep -Dtez.am.vertex.max-task-concurrency=1
> -Dtez.am.speculation.enabled=true -Dtez.task.timeout-ms=60000 -m 3 -mt 60000
> -ir 0 -irt 0 -r 0 -rt 0
> {noformat}
> 2. Let the 1st task run to completion and then stop the 2nd task so that a
> speculative attempt is scheduled. Once the speculative attempt is scheduled
> for the 2nd task, continue the original attempt and let it complete.
> {noformat}
> kill -STOP <pid>
> // wait a few seconds for a speculative attempt to kick off
> kill -CONT <pid>
> {noformat}
> 3. Kill the 3rd task, which will create a 2nd attempt
> {noformat}
> kill -9 <pid>
> {noformat}
> 4. The next thing to be drawn off of the queue will be the speculative
> attempt of the 2nd task. However, it is already completed, so it will just
> sit in the final state and the job will hang.
> Basically, for the failure to happen, the number of speculative tasks that
> are scheduled, but not yet ran has to be >= the concurrency of the job and
> there has to be at least 1 task failure.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
