Meng Zhu created MESOS-8459:
-------------------------------
Summary: Executor could linger without ever receiving any tasks
Key: MESOS-8459
URL: https://issues.apache.org/jira/browse/MESOS-8459
Project: Mesos
Issue Type: Bug
Components: executor
Reporter: Meng Zhu
An executor's initial tasks may be killed even after it has been registered. In
that case, the executor could linger forever.
In MESOS-8411, we have a short-term fix that checks an executor's completed and
terminated task queues to see if it had ever received any tasks. if the check
is false and there is no queued or launched tasks, agent will shutdown the
executor.
However, this check is not bullet-proof. The completedTasks queue is a
circular_buffer (current size 200) which means earlier completed tasks that are
possibly updated by the executor may be ejected and thus are missed by this
check. This would lead to false positive shutdowns.
Per discussion with [~vinodkone] and [~bmahler]. There are two long term
solutions.
The first one is to checkpoint additional executor states which indicates
whether the executor has ever received any tasks (no more inference from task
queue status);
The alternative is to add timeouts in the executor driver to shutdown lingering
executors automatically.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
