Ryan Williams created SPARK-11285:
-------------------------------------
Summary: Infinite TaskCommitDenied loop
Key: SPARK-11285
URL: https://issues.apache.org/jira/browse/SPARK-11285
Project: Spark
Issue Type: Bug
Components: Spark Core
Affects Versions: 1.5.0
Reporter: Ryan Williams
I've seen several apps enter this failing state in the last couple of days.
I've gathered all the documentation I can about two of them,
[application_1444948191538_0051|https://www.dropbox.com/sh/ku9btpsbwrizx9y/AAAXIY0VhMqFabJBCtTVYxtma?dl=0]
and
[application_1444948191538_0116|https://www.dropbox.com/home/spark/application_1444948191538_0116].
Both were run on Spark 1.5.0 on in yarn-client mode with dynamic allocation of
executors.
In application_1444948191538_0051, partitions 5808 and 6109 in stage-attempt
1.0 failed 7948 and 7921 times, respectively, before I killed the app. In both
cases, the first two attempts failed due to {{ExecutorLostFailure}}'s, and the
remaining ~7900 attempts all failed due to {{TaskCommitDenied}}'s, over ~6hrs
at a rate of about once per ~4s. See the last several thousand lines of
[application_1444948191538_0051/driver|https://www.dropbox.com/s/f3zghuzuxobyzem/driver?dl=0].
In application_1444948191538_0116, partition 10593 in stage-attempt 6.0 failed
its first attempt due to an ExecutorLostFailure, and then a subsequent 219
attempts in ~22mins due to {{TaskCommitDenied}}'s before I killed the app.
Again, [the driver logs|https://www.dropbox.com/s/ay1398p017qp712/driver?dl=0]
enumerate each attempt.
I'm guessing that the OutputCommitCoordinator is getting stuck due to early
failed attempts?
I'm trying to re-run some of these jobs on a 1.5.1 release and will let you
know if I repro it there as well.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
