First steps towards integration with WAS
The 0.19 release formally kicks off the alignment of Kueue with Kubernetes community initiatives centered around Workload-Aware Scheduling (WAS) and the newly unveiled Workload and PodGroup APIs. As an initial milestone, integration with the Kubernetes Scheduler Library has been delivered (see: SchedulerLibraryIntegration feature gate), ensuring that admission decisions in Kueue more accurately reflect standard Kubernetes scheduling semantics. The ultimate, ambitious objective is to establish full architectural synergy between these two projects.
Troubleshooting not admitted workloads
Both batch users and administrators gain a far more detailed understanding of why workloads remain queued, thanks to a new suite of features. When a job is awaiting admission, the system now provides precise, descriptive messages identifying specific reasons, such as ExceedsMaxQuota or TopologyPlacementFailed. To support this, new metrics — specifically kueue_unadmitted_workloads and kueue_local_queue_unadmitted_workloads — allow operators to track these occurrences, categorized cleanly by overall reason and granular cause.
Incremental dispatching improvements
The incremental dispatching mode in MultiKueue has been enhanced to adhere to the designated worker cluster sequence from the configuration. Additionally, it now provides more precise management over how many worker clusters are nominated during each phase of the dispatch process.
Feature maturity and more reasonable defaults
Among several experimental features that have transitioned in maturity to either beta or GA, I’d like to highlight AdmissionGatedBy, which is now enabled by default. To see an example of using it in practical scenarios, read the article on Automated Resource Configuration in Kubeflow.
The team has finally decided to enable WaitForPodsReady by default with a 30-minute timeout, ensuring that newcomers will benefit from out-of-the-box all-or-nothing scheduling. Speaking of out-of-the-box improvements, some of the default limits have also been raised: Kubernetes client limits are currently set at 300 QPS, while controller concurrency increases to 10 for Workloads and 5 for Local- and Cluster Queues. Additionally, the maximum number of PodSets per Workload has been increased to 18.
Stability, security and performance hardening
After applying the AI-assisted code analysis, the team resolved a record number of bugs related to Kueue’s stability, security and performance. The hardening effort continues, with the goal of resolving the remaining findings very soon.
Modernized Project Home
Finally, the official Kueue site has been redesigned. In addition to a fresh visual style, the community section now features an expanded, searchable directory of Kueue adopters, external articles, and videos. The documentation is also continuously being updated; for instance, see the recently introduced collection of vendor-specific Kueue materials.
For a comprehensive list of all updates and changes, you can explore the official changelog.
If you found this article helpful, follow me for more deep dives into Kueue and the evolving landscape of Kubernetes workload scheduling.
Kueue v0.19: What’s new? was originally published in Google Cloud – Community on Medium, where people are continuing the conversation by highlighting and responding to this story.
Source Credit: https://medium.com/google-cloud/kueue-v0-19-whats-new-93e8eb8e64c0?source=rss—-e52cf94d98af—4
About the Author
