Changes since v0.16.4:
Changes by Kind
Feature
- Helm: Add kueueViz.backend.ingress.enabled and kueueViz.frontend.ingress.enabled Helm values to allow disabling KueueViz ingress resources. (#10065, @david-gang)
- Helm: Add support to pass in custom issuerRef to allow for configuration of Issuers. (#10212, @MatteoFari)
- Introduced the
WorkloadNameShortenfeature gate to ensure generated Workload names do not exceed 63 characters. This prevents issues where Workload labels were invalid due to length. When enabled, the owner-based prefix is truncated to fit within the limit while maintaining uniqueness via a hash suffix. (#10129, @mbobrovskyi)
Bug or Regression
- FairSharing: Fix
FairSharingPrioritizeNonBorrowingto check per-flavor borrowing at every hierarchy level in hierarchical cohorts, not just at the ClusterQueue level. (#10203, @mukund-wayve) - RayJob integration: fix the autosaling scenarios when using ElasticJobsViaWorkloadSlices. In particular when
two consecutive scale ups happen. (#10160, @hiboyang) - Scheduling: fix the bug that scheduler could get stuck trying to preempt a workload due to the corruption of the
in-memory state tracking the pending preemptions (called preemptionExpectations). (#10206, @mimowo) - Strip managedFields from informer cache via DefaultTransform to reduce memory footprint on large clusters. (#10125, @jzhaojieh)
- TAS: Fix a bug where preemption with multiple resources sometimes fails (#10204, @mimowo)
- TAS: Fix nil pointer panic in TAS node reconciler when unadmitted workloads exist in the cluster. (#10039, @kannon92)
- TAS: Improved the performance of the node_controller Reconcile loop by introducing a new field indexer for Workloads. (#10051, @j-skiba)
- TAS: Workloads that require TAS but have a PodSet with a failed TAS request (e.g., more than one flavor assigned) are correctly rejected at admission with a clear Pending reason and message, rather than being admitted without TopologyAssignment. (#10228, @j-skiba)
- TAS: fix the bug that workloads which only specify resource limits, without requests, are not able to perform
the second-pass scheduling correctly, after Kueue restart, responsible for NodeHotSwap and ProvisioningRequests. (#10177, @mimowo)