Add rjgoyln's September contribution report for apache-yunikorn

[rjgoyln]

rjgoyln Monthly Contribution Report

• GitHub: https://github.com/rjgoyln
• Slack: https://opensource4you.slack.com/team/U0BMWP51QN5
• ASF JIRA: Dian-Xuan Yang
• Period: 2026-09-02 to 2026-10-05 (since my first YuniKorn contribution)

Summary

PR: 14 (12 merged, 2 in review, across all 6 YuniKorn repos), Code Review: 5, Issue Triage: 2 (plus 3 JIRAs filed)

YuniKorn committers merge by pushing the patch themselves, so merged PRs show as Closed on GitHub; the landed commits are on master (core, k8shim).

PR

Merged

• [YUNIKORN-3428] Fix fatal scheduler exit on concurrent taskMap access: fix an unlocked taskMap read racing with a new pod's write, which made the Go runtime kill the shim.
• [YUNIKORN-3427] Fix unsynchronised context and task access on deferred release: fix two races in flushReleaseableTasks, on the context's applications map and on the task's allocationKey.
• [YUNIKORN-3422] Publish the scheduler cache node lists atomically: publish the lazily built node lists atomically, since concurrent preemption predicates populated them under a read lock.
• [YUNIKORN-3358] Remove the unreachable replace-existing-ask branch: replace an unreachable branch that left sortedRequests inconsistent with an up-front duplicate-key rejection (in the 1.10.0 release branch).
• [YUNIKORN-3480] Fix the weekly e2e matrix so every K8s version runs: fix the weekly e2e, which had failed every week since 1.37 was added because of a missing comma and three nonexistent kind images.
• [YUNIKORN-3485] Remove unused UpdatePod from the shim KubeClient: remove dead code found in the YUNIKORN-3437 audit.
• [YUNIKORN-3484] Update GitHub Actions to latest versions (core, k8shim, scheduler-interface, web, release, site): bump the actions in all 6 repos, checking each major upgrade against the existing workflows and the ASF allowlist.

In Review

• [YUNIKORN-3481] Fix orphan allocation when a removal races with scheduling: fix an allocation left on its node until restart when its removal races with scheduling, a regression from YUNIKORN-2459 in 1.6.0.
• [YUNIKORN-3487] Cancel pod condition updates on shim shutdown: make pod condition updates cancellable so a slow API server no longer blocks shutdown.

Code Review

• [YUNIKORN-3479] Optimize preemption victim selection and trimming: found that the new victim ranking ignores priority ordering and scores nondeterministically, and suggested cached, key-sorted vectors that also cut the benchmark from 10.5 ms to 2.3 ms at n=1000.
• [YUNIKORN-3483] Reinstate Partition state transitions: raised a blocking issue where a queue removed from the config could never drain because its pending asks would stop being scheduled.
• [YUNIKORN-3478] Decouple ask queue quota accounting from victim sizes: pointed out that a flipped test assertion hid an over-preemption regression, which the author and I agreed to address in #1178.
• [YUNIKORN-3465] Add victimAllocationKeys in PreemptionPredicatesResponse: approved, suggesting the spec define victimAllocationKeys as an ordered subset of preemptAllocationKeys[0..index].
• [YUNIKORN-3421] Fix task release deadlocks in FSM callbacks: approved.

Issue Triage

• YUNIKORN-3437 Handle shutdown signal during K8s CRUD operations: audited the shim's 8 K8s write call sites for cancellability, mapped them to existing JIRAs, and split out YUNIKORN-3485 and YUNIKORN-3487.
• YUNIKORN-3231 ForeignAllocations ignored when pods use nodeSelector: reproduced it on 1.8.0 and traced it to informer event ordering, which YUNIKORN-3317 already fixed in 1.9.0.
• Filed YUNIKORN-3480, YUNIKORN-3485 and YUNIKORN-3487, all with PRs.

[rjgoyln]

cc @HuangTing-Yao @chenyulin0719

[henry3260]

Wow, that's a crazy amount of work! 🔥

[rjgoyln]

Wow, that's a crazy amount of work! 🔥

Thank you! handsome Chen ❤️