Saisai Shao created SPARK-13642:
-----------------------------------
Summary: Inconsistent finishing state between driver and AM
Key: SPARK-13642
URL: https://issues.apache.org/jira/browse/SPARK-13642
Project: Spark
Issue Type: Bug
Components: YARN
Affects Versions: 1.6.0
Reporter: Saisai Shao
Currently when running Spark on Yarn with yarn cluster mode, the default
application final state is "SUCCEED", if any exception is occurred, this final
state will be changed to "FAILED" and trigger the reattempt if possible.
This is OK in normal case, but there's a race condition when AM received a
signal (SIGTERM), no any exception is occurred. In this situation, shutdown
hook will be invoked and marked this application as finished with success, and
there's no another attempt.
In such situation, actually from Spark's aspect this application is failed and
need another attempt, but from Yarn's aspect the application is finished with
success.
This could happened in NM failure situation, the failure of NM will send
SIGTERM to AM, AM should make this attempt as failure and rerun again, not
invoke unregister.
So to increase the chance of this race condition, here is the reproduced code:
{code}
val sc = ...
Thread.sleep(30000L)
sc.parallelize(1 to 100).collect()
{code}
If the AM is failed in sleeping, there's no exception been thrown, so from
Yarn's point this application is finished successfully, but from Spark's point,
this application should be reattempted.
So basically, I think only after the finish of user class, we could mark this
application as "SUCCESS", otherwise, especially in the signal stopped scenario,
it would be better to mark as failed and try again (except explicitly KILL
command by yarn).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
