Greg Mann commented on MESOS-8576:

Still working on this one. The problem is that {{Docker::inspect()}} has retry 
logic embedded within the library function, since we often call it before a 
container has started running in order to detect that the container is up. So, 
to avoid repeatedly registering {{onDiscard}} callbacks with every retry (which 
would constitute a memory leak), we need to pass the "context" of the current 
{{docker inspect}} call through the async call chain, and also make it 
accessible to the {{onDiscard}} callback which we install onto the returned 
future. Since the Docker library is not currently a libprocess actor, this is a 
bit difficult.

WIP patch here: https://reviews.apache.org/r/65683/

> Improve discard handling of 'Docker::inspect()'
> -----------------------------------------------
>                 Key: MESOS-8576
>                 URL: https://issues.apache.org/jira/browse/MESOS-8576
>             Project: Mesos
>          Issue Type: Improvement
>          Components: containerization, docker
>    Affects Versions: 1.5.0
>            Reporter: Greg Mann
>            Assignee: Greg Mann
>            Priority: Major
>              Labels: mesosphere
> In the call path of {{Docker::inspect()}}, each continuation currently checks 
> if {{promise->future().hasDiscard()}}, where the {{promise}} is associated 
> with the output of the {{docker inspect}} call. However, if the call to 
> {{docker inspect}} becomes hung indefinitely, then continuations are never 
> invoked, and a subsequent discard of the returned {{Future}} will have no 
> effect. We should add proper {{onDiscard}} handling to that {{Future}} so 
> that appropriate cleanup is performed in such cases.

This message was sent by Atlassian JIRA

Reply via email to