Andreas Raster created MESOS-1845:
-------------------------------------
Summary: CommandInfo tasks may fail if scheduled after another
task with the same id has finished.
Key: MESOS-1845
URL: https://issues.apache.org/jira/browse/MESOS-1845
Project: Mesos
Issue Type: Bug
Reporter: Andreas Raster
I created a little test framework where I wanted to experiment with scheduling
tasks where running one task relies on the results of another, previously run
task. So in my test framework I would first schedule a task that would append
the string "foo" to a file, and after that one finishes I would schedule a task
that appends "bar" to the same file.
This worked well when using ExecutorInfo, but when I switched to using
CommandInfo instead (specifying commands like 'echo foo >> /share/foobar.txt'
in set_value()), it would most of the time fail in the second step when
attempting to append "bar". Occasionally, but very rarely, it would work though.
I couldn't find any meaningful log messages indicating what exactly went wrong.
The slave log would indicate that the tasks status changed to TASK_FAILED and
that that status update was sent correctly. The stdout log in the Sandbox would
indicate that the command 'exited with status 0'.
I could work around the issue when I specified task ids that were always
unique. Previously I would reuse the id of a previously run task, one that
appended "foo" to a file, after it finished in the followup task that would
append "bar" to a file.
It seems to me there might be something wrong when scheduling very short
running tasks with the same id quickly after each other.
Source code for my foobar framework:
http://paste.ubuntu.com/8459083
Build with:
g++ -std=c++0x -g -Wall foobar_framework.cpp -I. -L/usr/local/lib -lmesos -o
foobar-framework
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
