[
https://issues.apache.org/jira/browse/MESOS-3863?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15002673#comment-15002673
]
Joseph Wu commented on MESOS-3863:
----------------------------------
Inter-dependency between {{process_manager}} and {{socket_manager}} will
complicate things:
* {{process_manager}} holds the {{gc}} and various {{HttpProxy}} processes.
* {{socket_manager}} spawns {{HttpProxy}} processes and relies on {{gc}} to
clean them up.
* {{gc}} relies on {{socket_manager}} links to clean up processes.
{{process::finalize}} should:
# Clean up all processes other than {{gc}}. This will clear all links and
delete all {{HttpProxy}} s while {{socket_manager}} still exists.
# Close all sockets via {{SocketManager::close}}. All of {{socket_manager}} 's
state is cleaned up via {{SocketManager::close}}, including termination of
{{HttpProxy}} (termination is idempotent, meaning that killing {{HttpProxy}} s
via {{process_manager}} is safe).
# At this point, {{socket_manager}} should be empty and only the {{gc}} process
should be running. (Since we're finalizing, assume there are no threads trying
to spawn processes.) {{socket_manager}} can be deleted.
# {{gc}} can be deleted. This is currently a leaked pointer, so we'll also
need to track and delete that.
# {{process_manager}} should be devoid of processes, so we can proceed with
cleanup (join threads, stop the {{EventLoop}}, etc).
> Investigate the requirements of programmatically re-initializing libprocess
> ---------------------------------------------------------------------------
>
> Key: MESOS-3863
> URL: https://issues.apache.org/jira/browse/MESOS-3863
> Project: Mesos
> Issue Type: Task
> Components: libprocess, test
> Reporter: Joseph Wu
> Assignee: Joseph Wu
> Labels: mesosphere
>
> This issue is for investigating what needs to be added/changed in
> {{process::finalize}} such that {{process::initialize}} will start on a clean
> slate. Additional issues will be created once done. Also see [the parent
> issue|MESOS-3820].
> {{process::finalize}} should cover the following components:
> * {{__s__}} (the server socket)
> ** {{delete}} should be sufficient. This closes the socket and thereby
> prevents any further interaction from it.
> * {{process_manager}}
> ** Related prior work: [MESOS-3158]
> ** Cleans up the garbage collector, help, logging, profiler, statistics,
> route processes (including [this
> one|https://github.com/apache/mesos/blob/3bda55da1d0b580a1b7de43babfdc0d30fbc87ea/3rdparty/libprocess/src/process.cpp#L963],
> which currently leaks a pointer).
> ** Cleans up any other {{spawn}} 'd process.
> ** Manages the {{EventLoop}}.
> * {{Clock}}
> ** The goal here is to clear any timers so that nothing can deference
> {{process_manager}} while we're finalizing/finalized. It's probably not
> important to execute any remaining timers, since we're "shutting down"
> libprocess. This means:
> *** The clock should be {{paused}} and {{settled}} before the clean up of
> {{process_manager}}.
> *** Processes, which might interact with the {{Clock}}, should be cleaned up
> next.
> *** A new {{Clock::finalize}} method would then clear timers,
> process-specific clocks, and {{tick}} s; and then {{resume}} the clock.
> * {{__address__}} (the advertised IP and port)
> ** Needs to be cleared after {{process_manager}} has been cleaned up.
> Processes use this to communicate events. If cleared prematurely,
> {{TerminateEvents}} will not be sent correctly, leading to infinite waits.
> * {{socket_manager}}
> ** The idea here is to close all sockets and deallocate any existing
> {{HttpProxy}} or {{Encoder}} objects.
> ** All sockets are created via {{__s__}}, so cleaning up the server socket
> prior will prevent any new activity.
> * {{mime}}
> ** This is effectively a static map.
> ** It should be possible to statically initialize it.
> * Synchronization atomics {{initialized}} & {{initializing}}.
> ** Once cleanup is done, these should be reset.
> *Summary*:
> * Implement {{Clock::finalize}}. [MESOS-3882]
> * Implement {{~SocketManager}}.
> * Clean up {{mime}}.
> * Wrap everything up in {{process::finalize}}.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
