[
https://issues.apache.org/jira/browse/MESOS-9692?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16805465#comment-16805465
]
Meng Zhu commented on MESOS-9692:
---------------------------------
{noformat}
commit b65b57038cb495408078a343ce4401e12d329561
Author: Meng Zhu <[email protected]>
Date: Fri Mar 29 11:35:42 2019 -0700
Fixed a bug where disk quota may be under allocated.
The resources chopping logic in the allocator
currently overly shrinks resources with the same
name (e.g. vanilla disk and mount disk) if there
are multiple such resources. Currently only
different disk resources might share the same
name (we only chop unreserved/non-revocable/non-shared
resources). This would lead to disk quota being under
allocated. This patch corrects the shrinking logic.
See MESOS-9692.
Review: https://reviews.apache.org/r/70344
{noformat}
> Quota may be under allocated for disk resources.
> ------------------------------------------------
>
> Key: MESOS-9692
> URL: https://issues.apache.org/jira/browse/MESOS-9692
> Project: Mesos
> Issue Type: Bug
> Components: allocation
> Affects Versions: 1.5.3, 1.6.2, 1.7.2
> Reporter: Meng Zhu
> Assignee: Meng Zhu
> Priority: Major
> Labels: mesosphere, resource-management
>
> Due to a bug in the resources chopping logic:
> https://github.com/apache/mesos/blob/1915150c6a83cd95197e25a68a6adf9b3ef5fb11/src/master/allocator/mesos/hierarchical.cpp#L1665-L1668
> {noformat}
> if (Resources::shrink(&resource, limitScalar.get())) {
> targetScalarQuantites[resource.name()] -= limitScalar.get(); // bug
> result += std::move(resource);
> }
> {noformat}
> When chopping different resources with the same name (e.g. vanilla disk and
> mount disk), we only include one of the resources. For example, if a role has
> a quota of 100disk, and an agent has 50 vanilla disk and 50 mount disk, the
> offer will only contain 50 disk (either vanilla or the mount type). The
> correct behavior should be that both disks should be offered.
> Since today, only disk resources might have the same name but different
> meta-data (for unreserved/nonrevocable/nonshared resources -- we only chop
> this), this bug should only affect disk resources today.
> The correct code should be:
> {noformat}
> if (Resources::shrink(&resource, limitScalar.get())) {
> targetScalarQuantites[resource.name()] -= resource.scalar(); // Only
> subtract the shrunk resource scalar
> result += std::move(resource);
> }
> {noformat}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
