[
https://issues.apache.org/jira/browse/MESOS-2196?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14710124#comment-14710124
]
Lans Carstensen commented on MESOS-2196:
----------------------------------------
Has any more consideration been given to this?
For similar resource managers in the past we call this class of worker issue a
"cookie monster" - it results in workers taking tasks quickly and quickly
erroring them, breaking the tasks and flinging them around without actually
consuming any of them / doing useful work.
What's needed is a mechanism to signal that the worker is not in a "ready
state"[1] and can not receive Docker tasks.
[1] Best concise summary of concept I know is here:
http://robhirschfeld.com/2014/04/25/ready-state-infrastructure/
> Add option to Docker containerizer to check for adequate disk space prior to
> "docker pull" on slave
> ---------------------------------------------------------------------------------------------------
>
> Key: MESOS-2196
> URL: https://issues.apache.org/jira/browse/MESOS-2196
> Project: Mesos
> Issue Type: Improvement
> Components: docker
> Affects Versions: 0.20.1
> Reporter: Lans Carstensen
>
> This is a request for a flag similar in nature to
> https://reviews.apache.org/r/28190/ to ensure that there is adequate disk
> space on a mesos-slave prior to attempting to download a Docker image. This
> would avoid negatively impacting running tasks/containers and/or running a
> system out of disk space since this machine resource is currently unmanaged.
> If space is unavailable to run a particular task, It would be desirable to be
> able to invoke a preconfigured space reclamation workflow on the slave.
> Regarding implementation on the slave, one would have to query the /info
> endpoint to find storage information from the Docker storage driver. My
> particular use case is with devicemapper. One then has to resolve the list
> of image layers associated with a particular image tag, remove the list of
> layers already downloaded, and then query the registry for each layer's size
> to determine if all layers can be downloaded.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
