Skip to end of metadata
Go to start of metadata

You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 11 Next »

Scope

  • The functions for terminating task processes by the JobScheduler Master and Universal Agent have been extended to allow the use of both SIGTERM and SIGKILL signals on Unix servers:
    • SIGTERM is sent first and allows an orderly termination of task processes to take place within a limited period of time.
    • If the time allowed has been exceeded and the processes are still running then SIGKILL will be sent.
  • The information contained in this article draws together detailed information contained in a range of issues and should primarily be of interest to persons in engineering and to a lesser extent persons in operating functions.

Feature History

This feature has been implemented stepwise between releases 1.9.0 and 1.10.0 (see the table of issues below for more detailed information).

Issues

Support of this feature is subject to the following issues:

Key Summary T Created Updated Due Assignee Reporter P Status Resolution Fix Version/s
Loading...
Refresh

Use Case

Users who schedule programs and scripts that are aware of SIGTERM signals can implement clean-up procedures on receipt of the signal. Clean-up includes e.g. removal of temporary files, disconnect from a database and similar tasks.

  • The use of both SIGTERM and SIGKILL signals on Unix servers has the following advantages:
    • The use of SIGTERM before SIGKILL means that there is a greater chance of data being saved after the signal has been issued.

    • The SIGTERM signal can - in contrast with SIGKILL - be monitored, i.e. a pre-/post-processing Script can be carried out. This means that the ending of a task by the JobScheduler can be reacted to and the user process itself can be ended.

    • The implementation of SIGTERM allows post-processing methods such as spooler_process_after() to complete within the timeout period.
  • The time allowed between the SIGTERM and SIGKILL signals can be specified in the command using the timeout attribute (the default is 15 sec): <kill_task … timeout=".."/>

  • This feature can also be applied for:
    • remote processes, i.e. processes started by SSH and those started by an Agent,
    • child processes started by a process running on an agent (JS-1468).

Implementation

JobScheduler Master

The following operations can be carried out from the JobScheduler Operating Center interface (JOC) or by use of the command line:

  1. Operation: kill immediately
    • JOC sends <kill_task immediately="yes"/>
    • The process is killed immediately using the SIGKILL signal.
  2. Operation: terminate with timeout
    • JOC sends <kill_task immediately="yes" timeout="15"/>
    • The process receives a SIGTERM signal. Should that process not terminate within the specified timeout period then it will be killed with a SIGKILL signal.
  3. Operation: terminate
    • JOC sends <kill_task immediately="yes" timeout="never"/>
    • The process receives a SIGTERM signal. Monitoring of the process termination as described in Operation 2 above is not carried out.

JobScheduler Universal Agent (JUA)

The JobScheduler Universal Agent comes with two scripts - for Windows and for Unix systems - that allow parent and child tasks on the Agent to be killed.

The scripts are:

  • jobscheduler_agent_kill_task.cmd and
  • jobscheduler_agent_kill_task.sh

They are to be found  the ./bin folder of the JUA installation.

The script for the operating system being used is called with the parameter -kill-agent-task-id=...  and

  • finds the process containing the ID specified and
  • kills the process including all child processes.

The JUA start script starts the Agent with the new -kill-script parameter as follows:

  • by default the -kill-script is parameterized with the path to the kill script for Windows/Unix as appropriate,
  • the SCHEDULER_KILL_SCRIPT environment variable can be used to set a different kill script.

See  JS-1468 & JS-1495 for more detailed information.

Implementation Summary

The implementation of the different termination operations available for the JobScheduler Master and Universal Agent is summarized in the table below.

The termination operations available are:

  • Terminate: <kill_task immediately="true" timeout="...">
  • Kill: <kill_task immediately="true">
  • Timeout: a job with the timeout attribute. VERIFY
 Windows
Standalone
Linux
Standalone
Windows
Classic Agent
Linux
Classic Agent
Windows
Universal Agent
Linux
Universal Agent
Shell job      
Terminatenot supported(tick)not supported

(error) 2)

JS-1420 - Getting issue details... STATUS

not supported(tick)
Kill(tick)(tick)(tick)

(tick)

JS-1421 - Getting issue details... STATUS

(tick)

JS-1468 - Getting issue details... STATUS

(tick)

JS-1468 - Getting issue details... STATUS

Timeout

(tick)

JS-1463 - Getting issue details... STATUS

(tick)

JS-1463 - Getting issue details... STATUS

(tick)(tick) 

(tick)

JS-1468 - Getting issue details... STATUS

Shell job with monitor      
Terminatenot supported(tick)not supported(tick)not supported

(error) 3)

JS-1468 - Getting issue details... STATUS

Kill(tick)(tick)(tick)(tick)

(tick)

JS-1382 - Getting issue details... STATUS

(error) 1)

JS-1468 - Getting issue details... STATUS

Timeout(tick)(tick)   

(error) 1)

JS-1382 - Getting issue details... STATUS

API job      
Terminatenot supported(tick)not supported(tick)not supported(tick)
Kill(tick)(tick)(tick)(tick)(tick)(tick)
Timeout(tick)(tick)    
  1. The child process continues: the shell job (and Java process if applicable) is killed, but the child process (sleep or ping - see attached jobs) is detached from the process tree and continues to run.
  2. No effect: neither the shell job nor it's children receive a signal
  3. The Java process is terminated but it's child process (the shell job script) and child processes thereof (sleep command) do not receive a SIGTERM signal
  4. 1.9.2-1.9.4 the kill comes too late, after the task has ended normally

Delimitation

  • This feature is intended for Unix platforms that implement the SIGTERM and SIGKILL signals. It is not intended for Windows platforms for which exclusively the Kill Immediately command applies.
  • When using traps then please consider that the process created by the <shell> element receives the signal. Subsequent scripts that are called within the <shell> element will not receive the signal. You could therefore:
    • configure traps directly within the <shell> element. The shell process will then receive and handle the signal.
    • configure traps in a shell script that is added by an <include> element instead of being stated within the <shell> element. The included shell script will receive and handle the signal.
    • forward signals to subsequent shell scripts that are called within a <shell> element.
  • This feature has been fully implemented on the Universal Agent and It has been implemented for classic JobScheduler Agents (JS-1420).

Workarounds

Workaround Title ???

  • Should job scripts not be able to catch signals by traps then you can use a monitor script, i.e. a post-processing script, that would be called by JobScheduler on receipt of a SIGTERM signal (JS-1463). For example:

    Workaround for shell jobs with a timeout
    <job name="shell_with_javascript_monitor">
        <script  language="shell">
            <![CDATA[
    echo hello world!
    sleep 45
            ]]>
        </script>
    
        <monitor  name="process0" ordering="0">
            <script  language="java:javascript">
                <![CDATA[
    function spooler_process_before(){
    	return true;
    }
                ]]>
            </script>
        </monitor>
    
        <run_time />
    </job>

Examples

Example Title ..... ???

Download the Example

job_trap_sigterm.job.xml.zip

Description

This example contains a job that uses a SIGTERM trap to show the difference between the <kill_task> and <terminate_task> commands provided by JOC.

The job job_trap_sigterm.job.xml shows how to trap the terminate command provided by JOC.

  • Start the job
  • Terminate the task in JOC
  • You will see the log message: sigterm will be ignored
  • The task will continue
     

<?xml version="1.0" encoding="ISO-8859-1"?>
<job  title="test test">
    <script  language="shell">
        <![CDATA[
trap 'echo sigterm will be ignored' 15
for i in 1 2 3 4 5 6 7 8 9 0
do
date
sleep 10
done
sleep 60
        ]]>
    </script>

    <run_time />
</job>

References

 

  • No labels
Write a comment…