[
https://issues.apache.org/jira/browse/MESOS-6204?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15520140#comment-15520140
]
Jie Yu commented on MESOS-6204:
-------------------------------
commit 498d14e934233e4501597b43da3924bfe8b2de20
Author: Kevin Klues <[email protected]>
Date: Sat Sep 24 19:37:33 2016 -0700
Updated mesos containerizer to checkpoint container runtime information.
This includes checkpointing both the container pid and the status of
the container upon exit. This also includes an update to tests to
account for new 'init' process semantics in a container. That is, the
name of the init process of the container is now "mesos-containerizer"
not "sh".
Review: https://reviews.apache.org/r/51407/
> Introduce a "runtime" directory owned by the containerizer for checkpointing
> container information.
> ---------------------------------------------------------------------------------------------------
>
> Key: MESOS-6204
> URL: https://issues.apache.org/jira/browse/MESOS-6204
> Project: Mesos
> Issue Type: Task
> Reporter: Kevin Klues
> Assignee: Kevin Klues
> Fix For: 1.1.0
>
>
> Currently the containerizer optionally checkpoints the container pid into the
> agent meta directory if requested by the agent. Moveing forward, we would
> like to have a containerizer-specific runtime directory where the
> containerizer will checkpoint any container state it would like to maintain
> independent of the agent container information that it checkpoints.
> Currently, this will include the container pid as well as the container exit
> status. In conjunction with the changes to the launch helper introduced by
> MESOS-6088, the containerizer will now be able to better recover the
> container exit status if it happens to be offline at the time that a
> container finishes.
> This includes being able to reap the exit status from a checkpointed location
> in addition to reading it directly from the container as it exits.
> Originally we were planning on pushing some of this functionality down into
> the LinuxLauncher, but in the end we decided it made more sense to keep all
> checkpointing information in the containerizer itself. We will likely
> revisit this decision in the future.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
