LWKD: Week Ending September 22, 2024
Week: 2024-09-22
Developer News
You have one day (or less) left to vote for Steering Committee members.
The call for presentations for the Maintainer Summit in Kubecon India is now open. The Maintainer summit combines a Kubernetes Contributor Summit with contributor discussions and presentations by other CNCF projects.
Release Schedule
Next Deadline: Production Readiness, October 3
It's the second week of 1.32 and hopefully you're hard at work on your planned Enhancements.
KEP of the Week
KEP 2837: Pod level resource limits
Currently resource allocation in PodSpec is done at the container level. The scheduler aggregates the resources requested by all the containers to find a suitable Node for the Pod. The Pod API lacks a way to specify limits at the Pod level, limiting the flexibility and ease of resource management for Pods as a whole. This KEP extends the Pod API with a resource spec at the Pod level. This new feature can be used to complement the existing resource limits and make things easier for tightly coupled applications. The KEP explains how the resource limits will be applied in different cases when Pod level and container level requests and limits are specified, as well as how the OOM score calculation will be done.
This KEP is tracked for alpha stage in the upcoming v1.32 release.
Other Merges
- Prevented legacy allocator range misinitialization, preventing IP conflicts.
- Extended discovery GroupManager with Group lister interface
- Explicit control of metrics collection in
scheduler_perftests, supporting multi-namespace - Ensure
kubeadm join/resethandles etcd members only if their URLs/IDs are unique or exist - GPU tests using Jobs, simplifying the process to verify successful completion with
cupyinstead ofCUDAsamples - Make sure to trigger Node/Delete event
- Feature enhancement reinstating the Nvidia DaemonSet installation in the GCE test harness
- Feature(scheduler): more fine-grained Node QHint for nodeunschedulable plugin and fixes
- Optimized the
Unstructured.GetManagedFieldsfunction by eliminating unnecessary deep copying of JSON value - Register missing Pod event for
NodeUnschedulableplugin - Test improvements: nvidia GPU(s)
- Fix setting
resolvConfin drop-in kubelet config files - Make sure that the endpoints controller can reconcile the Endpoint object when it has more than 1000 addresses
- Added integration tests for
NodeUnschedulable,podtopologyspread&NodeResourcesFitin requeueing scenarios - Support added for API streaming
- Improvisation in precision of
Quantity.AsApproximateFloat64 - Adds an 8-length buffer to the
resourceupdates.Updatechannel to prevent blocking during device plugin data transmission to kubelet - If the application/json;as=Table content type is requested, the WatchList will respond with a 406 (Not Acceptable) error
- Improve the kubelet test coverage
- Prevent the garbage collector controller from blocking indefinitely on a cache sync failure
- Ensure that mismatched hostname labels and node names do not lead to incorrect pod scheduling or failures with
nodeAffinity - Test case added for parsing a WSL 2 kernel version
- Guarantee that restartable and non-restartable init containers are accounted
- Prevent Memory manager UnexpectedAdmissionError
spec.terminationGracePeriodSecondsshould not be overwritten byMaxPodGracePeriodSeconds
Promotions
Deprecated
- Remove obsolete test ClusterDns and fixes flaking
- Remove node general update event from EventsToRegister when QHint is enabled