----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/66984/ -----------------------------------------------------------
(Updated May 8, 2018, 9:36 a.m.) Review request for mesos, Benjamin Mahler, Kapil Arya, Till Toenshoff, and Vinod Kone. Changes ------- Patch updated: - Made exclusion set role aware; - Moved exclusion set to slave struct; - Added invariant check; - Added a todo in quota allocation stage; - Documented a design decision in the description. Repository: mesos Description (updated) ------- Schedulers that are below their fair share will continue to get allocated resources even if they don't need those resources. The expected behavior here is that schedulers will decline those resources for a long time (or forever). Not all schedulers do this, however, which means that some schedulers might get _starved_ of resources. Technically these schedulers are already at or above their fair share, but some operators feel that this is keeping the cluster underutilized. Rather than guarantee that all schedulers will decline with large timeouts this patch ensures that all other schedulers will see resources from the declined agent before they see resources from that agent again, even if there are now more resources available on that agent. To this end, each agent maintains an exclusion set of frameworks that declined its allocation earlier and skips them during the allocation. Note, a framework here is associated with each of its roles. If a framework declined resources allocated to one of its roles, it can still get allocations for its other roles. Note, an agent's exclusion set will be cleared if its attribute changes. Diffs (updated) ----- src/master/allocator/mesos/hierarchical.hpp 955ae3e6a9e3c790fb260311ec4b4ef725a826d3 src/master/allocator/mesos/hierarchical.cpp 1000968be6a2935a4cac571414d7f06d7df7acf0 Diff: https://reviews.apache.org/r/66984/diff/4/ Changes: https://reviews.apache.org/r/66984/diff/3-4/ Testing ------- make check Dedicated test in #66994 Thanks, Meng Zhu
