[jira] [Commented] (MESOS-1654) Expose ephemeral ports per container as a resource

["Jie Yu (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MESOS-1654?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14080416#comment-14080416 ] 

Jie Yu commented on MESOS-1654:
-------------------------------

One question we need to address here is: some resources are tied with the underlying isolation strategy, how should we handle them in general. For example, if we use a mac vlan based network isolator (assuming we have enough IPs), ephemeral ports resource does not seem to make much sense since each container has a separate network namespace, and can use all the available ephemeral ports. Take a step further, the existing 'ports' resource may not make sense once we can assign each container a separate IP since two tasks in two containers are allow to listen to the same port.

[~tweingartner] suggested to a solution: each slave, when registering, specify the maximum number of containers it can support. The master keeps track of that and does not send offer to the frameworks if that limit has been reached. To me, this is a generic way of dealing with the 'limitation' imposed by the underlying isolation strategy.

Any comments?

> Expose ephemeral ports per container as a resource
> --------------------------------------------------
>
>                 Key: MESOS-1654
>                 URL: https://issues.apache.org/jira/browse/MESOS-1654
>             Project: Mesos
>          Issue Type: Story
>    Affects Versions: 0.20.0
>            Reporter: Dominic Hamon
>            Priority: Critical
>
> Ephemeral port static allocation requires fixed limit of containers per host. This is not exposed to frameworks so they will receive TASK_LOST notifications without the ability to proactively control this or react accordingly.
> By exposing the ephemeral port ranges assigned to containers as a resource, frameworks can potentially fit more containers per host, but also won't receive offers for slaves that can't run any more tasks due to hitting the limit.



--
This message was sent by Atlassian JIRA
(v6.2#6252)