> On Feb. 12, 2019, 7:34 p.m., Benjamin Bannier wrote:
> > src/tests/hierarchical_allocator_tests.cpp
> > Lines 2924-2929 (patched)
> > <https://reviews.apache.org/r/69942/diff/1/?file=2124739#file2124739line2924>
> >
> > This test actually is more than a regression test for MESOS-9554 as it
> > validates to some extent that framework allocation decisions are
> > independent (but for possibly decreasing the available resource pool).
> >
> > Maybe rephrase along the lines of
> > ```
> > This test validates with one case that framework allocation decisions
> > are independent (but for possibly decreasing the available resource pool).
> > In particular, we validate that the decision not to allocate resources of
> > an agent to a particular framework does not influence to allocate the same
> > resources to another agent. In this test we make agent resources not
> > allocatable to a framework by refining the resource reservations while the
> > framework which should be allocated these resources based on DRF share is
> > not capable of receiving them (it lacks the `RESERVATION_REFINEMENT`
> > capability).
> >
> > This is a regression test for MESOS-9554.
> > ```
> >
> > We could also rename the test, e.g., to
> > `HierarchicalAllocatorTest.AllocationIndependent` (doesn't read to well,
> > either). This tests models a whole class of scenarios we currently don't
> > validate.
>
> Benjamin Mahler wrote:
> Thanks for the suggestion. I will have to think on this, and so I'll
> leave it open for now
Thinking more about this, "allocations have no effect on each other" doesn't
seem right. A better invariant might that absent quota all cluster resources
get allocated if at least one framework could receive them. We could even
generalize this
Given
- `k` reasons resources could not be allocatable to a framework (e.g., filters
for whole agents or parts of them, framework capabilities)
- `n > k` frameworks
For all assignments of reason to `n - 1` frameworks (reasons can be repeated;
an assignement doesn't need to use all `k` reasons, `n`th framework can always
receive any resource):
For all DRF sortings of frameworks:
sum of allocations == total resources
If we care about this, it should be possible to generate all possible
configurations and check them.
There might be other invariants which could be checked in a similar way. Quota
be treated similarly, but we'd need to include quota headroom.
- Benjamin
-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/69942/#review212751
-----------------------------------------------------------
On Feb. 11, 2019, 7:21 p.m., Benjamin Mahler wrote:
>
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/69942/
> -----------------------------------------------------------
>
> (Updated Feb. 11, 2019, 7:21 p.m.)
>
>
> Review request for mesos, Benjamin Bannier and Meng Zhu.
>
>
> Bugs: MESOS-9554
> https://issues.apache.org/jira/browse/MESOS-9554
>
>
> Repository: mesos
>
>
> Description
> -------
>
> This test ensures that the framework allocation loop does not
> incorrectly break in the case that a framework is incapable
> of receiving the resources on an agent. In this case, the loop
> should continue as other frameworks may be capable of receiving
> the resources.
>
>
> Diffs
> -----
>
> src/tests/hierarchical_allocator_tests.cpp
> cc88afbad1b4e6bf707cb13b50c964aa01f9a3ee
>
>
> Diff: https://reviews.apache.org/r/69942/diff/1/
>
>
> Testing
> -------
>
> Ensured it fails on master, succeeds after fix to MESOS-9554.
>
> Ran in repetition.
>
>
> Thanks,
>
> Benjamin Mahler
>
>
