[
https://issues.apache.org/jira/browse/MESOS-8840?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16470555#comment-16470555
]
Greg Mann commented on MESOS-8840:
----------------------------------
cc [~abudnik]
> `cpu.cfs_quota_us` may be accidentally set for command task using docker
> during agent recovery.
> -----------------------------------------------------------------------------------------------
>
> Key: MESOS-8840
> URL: https://issues.apache.org/jira/browse/MESOS-8840
> Project: Mesos
> Issue Type: Bug
> Components: containerization
> Affects Versions: 1.2.0, 1.2.1, 1.2.2, 1.2.3
> Reporter: Meng Zhu
> Priority: Critical
>
> Prior to Mesos 1.3, docker containerizer does not honor the flag
> `–cgroups_enable_cfs` for command task when creating the container, a patch
> ported this flag to docker command executor only up to 1.3 (MESOS-6134)
> However, docker containerizer honors the flag when updating containers:
> https://github.com/apache/mesos/blob/7559c9352c78912526820f6222ed2b17ad3b19cf/src/slave/containerizer/docker.cpp#L1726
> For non-command tasks, docker containerizer always `update` the resources
> during launch:
> https://github.com/apache/mesos/blob/7559c9352c78912526820f6222ed2b17ad3b19cf/src/slave/containerizer/docker.cpp#L1325-L1330
> For command tasks, it is not the case:
> https://github.com/apache/mesos/blob/7559c9352c78912526820f6222ed2b17ad3b19cf/src/slave/containerizer/docker.cpp#L1271-L1277
> However, when recovering the executor, `update` is called for both command
> and non-command tasks.
> This means, for command task, the cpu cgroup cfs settings would change when a
> command executor is recovered. Specifically, recovered command executors will
> have cfs set while all other command executors will not. This may lead to a
> drastic change in the container resource usage depending on the system load.
> To maintain backward compatibility, we probably want to avoid setting the
> `cpu.cfs_quota_us` field in `update` if the field is not already set.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
