Changes since v0.15.1:
Changes by Kind
Feature
- Ray: Support RayJob InTreeAutoscaling by using the ElasticJobsViaWorkloadSlices feature. (#8284, @hiboyang)
Bug or Regression
-
Kubeflow TrainJob v2: fix the bug to prevent duplicate pod template overrides when starting the Job is retried. (#8271, @j-skiba)
-
MultiKueue: Fixed status sync for CRD-based jobs (JobSet, Kubeflow, Ray, etc.) that was blocked while the local job was suspended. (#8344, @IrvingMg)
-
MultiKueue: fix the bug that for Pod integration the AdmissionCheck status would be kept Pending indefinitely,
even when the Pods are already running.The analogous fix is also done for the batch/Job when the MultiKueueBatchJobWithManagedBy feature gate is disabled. (#8288, @IrvingMg)
-
Scheduling: fix a bug that evictions submitted by scheduler (preemptions and eviction due to TAS NodeHotSwap failing)
could result in conflict in case of concurrent workload modification by another controller.
This could lead to indefinite failing requests sent by scheduler in some scenarios when eviction is initiated by
TAS NodeHotSwap. (#8313, @mbobrovskyi) -
TAS NodeHotSwap: fixed the bug that allows workload to requeue by scheduler even if already deleted on TAS NodeHotSwap eviction. (#8310, @mbobrovskyi)
-
TAS: fix a performance bug that continues reconciles of TAS ResourceFlavor (and related ClusterQueues)
were triggered by updates to Nodes' heartbeat times. (#8355, @PBundyra) -
TAS: fixed performance issue due to unncessary (empty) request by TopologyUngater (#8333, @mbobrovskyi)
Other (Cleanup or Flake)
- Improve error messages for validation errors regarding WorkloadPriorityClass changes in workloads. (#8352, @olekzabl)
- MultiKueue: improve the MultiKueueCluster reconciler to skip attempting to reconcile and throw errors
when the corresponding Secret or ClusterProfile objects don't exist. The reconcile will be triggered on
creation of the objects. (#8290, @mszadkow)