Dale Richardson created YUNIKORN-3384:
-----------------------------------------
Summary: Partition accounting property fuzzer
Key: YUNIKORN-3384
URL: https://issues.apache.org/jira/browse/YUNIKORN-3384
Project: Apache YuniKorn
Issue Type: Test
Components: core - scheduler, shim - kubernetes
Reporter: Dale Richardson
A property fuzzer for the core's partition-level accounting. The shim cache
fuzzer
(YUNIKORN-3373) guards shim bookkeeping; nothing guarded the core's. This is the
structural guard for that class from the core side.
h3. Design
Seeded, weighted random operation sequences (several hundred ops per seed,
several
seeds) against a real PartitionContext with a small queue hierarchy, driven
through the same entry points the RM uses: add/remove node (removal with live
allocations included), add/remove app, add ask, TryAllocate cycles, release with
each TerminationType, replace-ask, reservation create/cancel, required-node
asks.
After every operation it checks:
* Q1 each queue's allocated == sum of its children (recursive)
* Q2 each leaf queue's allocated == sum of its apps' allocated
* Q3 usage never negative, never over a set maximum
* Q4 user/group tracked usage == sum of owned apps' allocated
* Q5 each node's allocated == sum of allocations on it
* Q6 queue pending == sum of app pending asks
* R1 every reservation references a live app + ask + node
plus a running allocated-minus-released model cross-check.
h3. Verification
Green on master across all seeds (this is a guard, not necessarily a finder).
Its
teeth are proven by mutation testing: five deliberate accounting breaks each
fire
the expected invariant (e.g. dropping queue.DecAllocatedResource fires
Q2/Q3/model;
dropping the parent recursion fires Q1; breaking the user tracker fires Q4). All
mutations reverted -- the committed diff is test-only. Runtime ~4s.
h3. Findings
Two real core defects were found while building this and are filed separately:
YUNIKORN-3379 (PLACEHOLDER_REPLACED crash) and YUNIKORN-3380
(RemoveAllAllocations
quota leak). Their trigger operations are deliberately excluded from the
committed
fuzzer so it stays green and mergeable; a follow-up can add them once the bugs
are
fixed.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
