[
https://issues.apache.org/jira/browse/SPARK-2881?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14088018#comment-14088018
]
Mridul Muralidharan edited comment on SPARK-2881 at 8/6/14 6:45 PM:
To add, this will affect spark whenever tmp directory is not overridden via
java.io.tmpdir to something ephemeral.
So local, yarn client, standalone should be affected by default (unless I
missed something in the scripts).
I am not very of how mesos runs jobs, so cant comment about that, anyone care
to add ?
A workaround I can think of is to always set 'org.xerial.snappy.tempdir' to a
randomly generated directory under java.io.tmpdir as part of spark startup
(only) once : which will cause snappy to use that directory and avoid this
issue.
Since snappy is the default codec now, I am +1 on marking this as a blocker for
release
was (Author: mridulm80):
To add, this will affect spark whenever tmp directory is not overridden via
java.io.tmpdir to something ephemeral.
So local, yarn client, standalone should be affected by default (unless I
missed something in the scripts).
I am not very of how mesos runs jobs, so cant comment about that, anyone care
to add ?
A workaround I can think of is to always set 'org.xerial.snappy.tempdir' to a
randomly generated directory under java.io.tmpdir as part of spark startup
(only) once : which will cause snappy to use that directory and avoid this
issue.
Since snappy is the default codec now, I am marking this as a blocker for
release
Snappy is now default codec - could lead to conflicts since uses /tmp
-
Key: SPARK-2881
URL: https://issues.apache.org/jira/browse/SPARK-2881
Project: Spark
Issue Type: Bug
Components: Spark Core
Affects Versions: 1.1.0
Reporter: Thomas Graves
Priority: Blocker
I was using spark master branch and I ran into an issue with Snappy since its
now the default codec for shuffle.
The issue was that someone else had run with snappy and it created
/tmp/snappy-*.so but it had restrictive permissions so I was not able to use
it or remove it. This caused my spark job to not start.
I was running in yarn client mode at the time. Yarn cluster mode shouldn't
have this issue since we change the java.io.tmpdir.
I assume this would also affect standalone mode.
I'm not sure if this is a true blocker but wanted to file it as one at first
and let us decide.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
-
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
