$ the-wire · showcase
DRA deadlock fix lands, watch cache tightens List selectors
By RepoJournal · Filed · About Kubernetes · Composed from the cited sources · methodology
10 people shipped this
Kubernetes fixed a DRA ResourceSlice controller deadlock that could stall publication entirely, while serathius' watch cache correctness work moved from unit verification in kubernetes/kubernetes to published apiserver commits.
DRA: avoid deadlocks in ResourceSlice controller Update and kubelet plugin shutdown kubernetes/kubernetes
With ReconcilePoolWithName set, Controller.Update rejects resources containing other pools and reports it through ErrorHandler while still holding c.mutex; a handler that replaces slices by calling Update blocks on that same mutex, so if previous resources had pools the worker blocks too and Stop never returns. The patch removes that deadlock path in the ResourceSlice controller and kubelet plu...
Implement List label selector verification kubernetes/kubernetes
The watch cache now verifies that List responses actually match the label selector they were served for, rather than trusting the index path. The merged apiserver commit makes this the correctness baseline that the rest of the watch cache work builds on.
Make snapshotter aware of resourceVersion and enabled kubernetes/kubernetes
The snapshotter gains awareness of resourceVersion and of whether snapshotting is enabled, with any enable/disable transition requiring an explicit reset so callers cannot forget one. This is groundwork for returning RV from storage through GetExactSnapshot in a self-consistent way.
Add GangPreemptionHeroWorkload benchmark to scheduler_perf kubernetes/kubernetes
wojtek-t added a GangPreemptionHeroWorkload benchmark to scheduler_perf as the baseline for upcoming workload-aware preemption optimizations. On the 4900-node test, preemption evaluation (PodGroupPostFilter) took about 129.13s, selecting 1,490 victims and reprieving the remaining 1,430 candidates, against a 10.22s initial workload pass.
Merge pull request #142607 from cwedgwood/fix/max-header-value-count kubernetes/apiserver
The apiserver now accepts large identity header sets under Go 1.27, raising the previous max header value count limit. If you front the API server with an authenticating proxy that emits many identity headers, those requests no longer get rejected.
apimachinery: add MinimumQuantity and MaximumQuantity validators kubernetes/apimachinery
apimachinery added MinimumQuantity and MaximumQuantity validators, with Cmp staying exact where Value() saturates at the int64 limits, and MinError/MaxError accepting the bound as a string since a Quantity does not format with %v. The matching code-generator work lets +k8s:minimum and +k8s:maximum carry resource.Quantity payloads as integers or quoted strings like +k8s:minimum="1Gi", parsed at ...
