[
https://issues.apache.org/jira/browse/AURORA-1056?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14290440#comment-14290440
]
Kevin Sweeney commented on AURORA-1056:
---------------------------------------
AURORA-1014 is related and IMO the better way forward here. We should be
pushing our users to avoid mutable docker images, instead resolving them to
stable identifiers at job create/update time so that rolling upgrades and
rollbacks work as expected.
> always send force_pull_image in ContainerInfo
> ---------------------------------------------
>
> Key: AURORA-1056
> URL: https://issues.apache.org/jira/browse/AURORA-1056
> Project: Aurora
> Issue Type: Bug
> Reporter: Jay Buffington
>
> mesos 0.21 introduced "force_pull_image" which IMHO, is a terrible idea and
> changed existing behavior.
> In 0.20 the image you specify will always be "docker pull"ed which means that
> if there was a newer version in the registry than the one in the cache you
> would get it.
> After 0.21 force_pull_image defaults to false, so unless you make a code
> change the "docker pull" doesn't happen. This means that if your cache has
> an older copy of the image, you run whatever happens to be in the cache.
> This results in starting 5 instances of the job, and 3 might one the old
> version (because it was cached on that slave) and 2 might one the new version
> (because those two slaves didn't have it cached), which is never what users
> want.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
