Dale Richardson created YUNIKORN-3399:
-----------------------------------------
Summary: Align RollbackAllocation tests with the ask-presence
guard from YUNIKORN-3360
Key: YUNIKORN-3399
URL: https://issues.apache.org/jira/browse/YUNIKORN-3399
Project: Apache YuniKorn
Issue Type: Bug
Components: core - scheduler
Reporter: Dale Richardson
h3. Problem
{{make test}} fails on master with two failures in
{{{}pkg/scheduler/objects{}}}:
* {{TestRollbackAllocationAskNotTracked}} — {{failed to locate ask with key
alloc-1 for rollback: rollback of the confirmed allocation should have
succeeded}}
* {{TestApplicationPropertyFuzzHistogram}} — {{totalGhostRollbacks is <= 0: no
seed ever rolled back an ask missing from sa.requests}}
Both fail deterministically on a clean master checkout; nothing else in the
package fails.
h3. Cause
A semantic conflict between two independently merged PRs. There was no textual
conflict, so both merged cleanly:
* YUNIKORN-3360 (commit 52aca85, merged 2026-08-18) added a guard to
{{{}RollbackAllocation{}}}: the rollback is now rejected when {{sa.requests}}
no longer holds the ask.
* YUNIKORN-3352 (commit 2a271ec, merged 2026-08-20) added tests written
against the pre-3360 behaviour, where rolling back such a "ghost" allocation
(ask already dropped from {{{}sa.requests{}}}, allocation still in
{{sa.allocations}} awaiting shim confirmation) succeeded:
** {{TestRollbackAllocationAskNotTracked}} asserts the ghost rollback succeeds
and leaves the pending histogram untouched.
** {{TestApplicationPropertyFuzzHistogram}} asserts at least one fuzz seed
performed a successful ghost rollback, as a coverage guarantee for that path.
With the guard in place a ghost rollback always returns an error, so both
assertions can no longer hold.
h3. Fix
Test-only change; pins the YUNIKORN-3360 semantics:
* {{{}TestRollbackAllocationAskNotTracked{}}}: assert the rollback is rejected
and that application state (pending histogram, askMaxPriority, pending
resource, allocations) is unchanged.
* {{{}TestApplicationPropertyFuzzHistogram{}}}: the fuzzer's reference model
already treats a rollback error as a no-op, so only the coverage assertion
changes — count ghost rollbacks that were attempted and rejected rather than
ones that succeeded.
h3. Note
The YUNIKORN-3360 guard was added on the assumption the ask is always present.
The ghost state is reachable: {{removeAsksInternal("")}} wipes {{sa.requests}}
while {{sa.allocations}} survives until the shim confirms the releases, and a
{{SCHEDULING_FAILED_ON_RM}} release arriving in that window reaches
{{{}RollbackAllocation{}}}. With the guard, that release is now dropped without
removing the allocation from the application, node or queue. That question is
independent of the test breakage and will be raised on YUNIKORN-3360; this
issue only fixes the tests.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
