[
https://issues.apache.org/jira/browse/MESOS-5295?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15354604#comment-15354604
]
Qian Zhang commented on MESOS-5295:
-----------------------------------
RR: https://reviews.apache.org/r/49354/
> The task launched by non-checkpointed HTTP command executor will keep running
> till executor shutdown grace period (5s) after agent process exits.
> -------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: MESOS-5295
> URL: https://issues.apache.org/jira/browse/MESOS-5295
> Project: Mesos
> Issue Type: Bug
> Components: HTTP API
> Reporter: Qian Zhang
> Assignee: Qian Zhang
>
> When I test HTTP command executor, I found an issue, here is my steps:
> 1. A framework which has no checkpoint enabled launches a long running task
> (e.g., sleep 1000).
> 2. After the task is running, kill the agent.
>
> Then I see the HTTP command executor will terminate after 5s
> ("DEFAULT_EXECUTOR_SHUTDOWN_GRACE_PERIOD"), but the task will always run.
> This behavior is not consistent with driver based command executor: after
> agent is killed, that executor will kill the task and then self terminate
> after 1s (there is a "os::sleep(Seconds(1));" in "reaped()").
> The root cause of this difference is, for driver based command executor, when
> the driver found agent is down, it will call executor->shutdown()
> (https://github.com/apache/mesos/blob/0.28.1/src/exec/exec.cpp#L487), so the
> executor will kill the task and then self terminate. But for HTTP command
> executor, its "disconnected()" will be called
> (https://github.com/apache/mesos/blob/0.28.1/src/executor/executor.cpp#L388)
> when agent is down, and currently we do not do anything in its
> "disconnected()", so the task will keep running and the executor will be
> killed after 5s
> (https://github.com/apache/mesos/blob/0.28.1/src/executor/executor.cpp#L623).
> The behavior of driver based command executor is correct, we need to make
> sure HTTP command executor kills the task when agent is down if checkpoint is
> not enabled.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
