[
https://issues.apache.org/jira/browse/MESOS-35?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13155628#comment-13155628
]
Andy Konwinski commented on MESOS-35:
-------------------------------------
I believe this issue no longer manifests (as of SVN r1205259) if you don't
specify a value for work_dir when running the mesos slave (i.e. if you let the
default values be used for them, which is either mesos_home or ".").
Michael and I use deploylib (deploylib.com), in which we run mesos on EC2 and
we set it up to use a config file in which we specify that work_dir = /mnt and
this causes the links (in the slave webui) for a framework's executors' stderr
and stdout to break. The erro message says that the /mnt/slave/... directory
can't be found. When I manually look, the file that is looking for is in
/mnt/work/slave/... (notice the extra "/work").
It appears that with these config settings, slave.cpp puts files in a different
directory (relative to work_dir from conf) than webui/slave/webui.py looks for
them in.
In src/slave/slave.cpp, when we're building up the directory where the
executors' stderr and stdout will eventually go, if work_dir is present as a
conf value, we should not append "/work" to it, but we currently do. See
snippet below, starting at line 1381:
<code>
1381 string workDir = ".";
1382 if (conf.contains("work_dir")) {
1383 workDir = conf.get("work_dir", workDir);
1384 } else if (conf.contains("home")) {
1385 workDir = conf.get("home", workDir);
1386 }
1387
1388 workDir = workDir + "/work";
1389
1390 std::ostringstream out(std::ios_base::app | std::ios_base::out);
1391 out << workDir << "/slaves/" << id
1392 << "/frameworks/" << frameworkId
1393 << "/executors/" << executorId;
</code>
... having that extra "/work" in the path is breaking things, since when
work_dir is set, src/slave/webui.cpp doesn't append "/work" to it (see starting
at line 80):
<code>
80 webuiPort = conf.get("webui_port", "8081");
81 logDir = conf.get("log_dir", FLAGS_log_dir);
82 if (conf.contains("work_dir")) {
83 workDir = conf.get("work_dir", "");
84 } else if (conf.contains("home")) {
85 workDir = conf.get("home", "") + "/work";
86 } else {
87 workDir = "work";
88 }
</code>
I'll attach a small patch that doesn't append "/work" to workDir in slave.cpp
if workDir is set from the "work_dir" conf flag.
> Slave webui needs to correctly display stdout/stderr of executors.
> ------------------------------------------------------------------
>
> Key: MESOS-35
> URL: https://issues.apache.org/jira/browse/MESOS-35
> Project: Mesos
> Issue Type: Bug
> Components: slave, webui
> Reporter: Benjamin Hindman
>
> There have been some changes to the work directories for frameworks due to
> multi-executor support and so the slave webui is no longer properly linking
> to executors stdout/stderr.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
