[
https://issues.apache.org/jira/browse/MESOS-3341?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15097513#comment-15097513
]
Jessica Hartog commented on MESOS-3341:
---------------------------------------
After lengthy discussion with people at the MesosCon Hackathon earlier this
year (where the idea for this ticket was fleshed out), I'm of the opinion that
out of resolution tasks should go through a deprecation cycle similar to the
one proposed for MESOS-1807. The reason for this is because rounding (with or
without a warning) will not be communicated to the guest framework's accounting
mechanism. Right now there is no mechanism for reporting that a framework
received a different amount of CPU/memory than what was requested, and even if
this were to be introduced, it would require additional logic changes inside
each guest framework to be able to account for this situation. Task error is
already in place and does not require changes to every guest framework for the
accounting on the mesos-side and the accounting on the framework-side to match.
> Introduce Resource Resolution
> -----------------------------
>
> Key: MESOS-3341
> URL: https://issues.apache.org/jira/browse/MESOS-3341
> Project: Mesos
> Issue Type: Improvement
> Reporter: Jessica Hartog
> Priority: Minor
>
> After MESOS-1807, Mesos containers require >= 0.01 CPU resources and >= 32MB
> Memory resources. In order to simplify accounting, Mesos should introduce
> resource resolution.
> For example, it is possible to launch a task with 1.0000000000001 CPU (as it
> exceeds the minimum number of CPUs and is therefore considered valid). The
> fractional component of this task does not benefit the running process, and
> can introduce floating point errors when Mesos is accounting its offers
> (which we have already seen happening in MESOS-1867 and MESOS-2635). A
> solution to this could be disallowing tasks with finer granularity than the
> required resolution.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
