[ https://issues.apache.org/jira/browse/MESOS-2176?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Benjamin Mahler updated MESOS-2176: ----------------------------------- Sprint: Twitter Mesos Q4 Sprint 5 > Hierarchical allocator inconsistently accounts for reserved resources. > ----------------------------------------------------------------------- > > Key: MESOS-2176 > URL: https://issues.apache.org/jira/browse/MESOS-2176 > Project: Mesos > Issue Type: Bug > Components: allocation > Reporter: Benjamin Mahler > Assignee: Benjamin Mahler > > Looking through the allocator code for MESOS-2099, I see an issue with > respect to accounting reserved resources in the sorters: > Within {{HierarchicalAllocatorProcess::allocate}}, only unreserved resources > are accounted for in the sorters, whereas everywhere else (add/remove > framework, add/remove slave) we account for both reserved and unreserved. > From git blame, it looks like this issue was introduced over a long course of > refactoring and fixes to the allocator. My guess is that this was never > caught due to the lack of unit-testability of the allocator (unnecessarily > requires a master PID to use an allocator). > From my understanding, the two levels of the hierarchical sorter should have > the following semantics: > # Level 1 sorts across roles. Only unreserved resources are shared across > roles, and therefore the "role sorter" for level 1 should only account for > the unreserved resource pool. > # Level 2 sorts across frameworks, within a role. Both unreserved and > reserved resources are shared across frameworks within a role, and therefore > the "framework sorters" for level 2 should each account for the reserved > resource pool for the role, as well as the unreserved resources _allocated_ > inside the role. -- This message was sent by Atlassian JIRA (v6.3.4#6332)