This Thursday is the May Community Meeting at 10am PDT/1700 UTC. SIGs Testing, UI, API Machinery, and Usability are presenting.
WG LTS wants to mark a year of patch support “implementable” based on the manifest fact that we are already supporting 1.16 for a year.
Next Deadline: Enhancements Freeze, May 19th
Is your enhancement issue filed? Your KEP marked “implementable” and complete? It’s now, or wait for 1.20 (or file an exception).
All supported versions (1.16, 1.17, and 1.18) will be getting updated on Wednesday
Kubernetes supports and requires klog v2 in v1.19, so get ready to update/rewrite your log-watching tools when you eventually upgrade. This change to logging is the major step in implementing structured logging, which will allow folks to impement better log digestion tools and do less ad-hoc text parsing. This was also joined by a new feature in the dependency linter to ensure that klog v1 (or future libraries) do not accidentally be added back in.
A good reminder about race conditions between controllers, using Update or UpdateStatus calls in places where multiple controllers write to the same structure can lead to either failed writes or stomping on updates if you aren’t careful. Using a Patch request instead allows for more careful control over this, and in the future Server Side Apply will potentially allow for more direct detection of conflicts through field tracking.
#91046: kubelet no longer registers “beta.kubernetes.io/os” and “beta.kubernetes.io/arch” node labels to apiserver
beta.kubernetes.io/arch automatic node labels were deprecated back in 1.14 and have now been removed entirely. Make sure to check all your node selector and affinity settings to ensure they are using the
- Kubelet metrics work on Windows nodes with init-containers present
- Don’t schedule “first pods” on nodes with the wrong topology just because they have affinity, and move unschedulable pods to the queue synchronously so that they don’t create a race condition
- Prevent replacing pods during HA failover events from corrupting the scheduler cache
- kubeadm checks
/readyzendpoints separately instead of relying on
- Keep PVC size requests from overflowing INT64
- CRI-O is faking some inapplicable memory metrics so that it works with HPA
- Cinder Volumes are Ext4 by default
- The Scheduler Framework adds configurable CPU and Memory weighting to its
- The CRI API lets you annotate Image Specs, mostly to support Windows
kubelet_certificate_manager_client_ttl_secondsreplaces the non-functional
beta.kubernetes.io/archnode labels are deprected, to be removed in 1.22 (back-ported to all versions)
--chaos-chancetesting flags have been removed from the kubelet, which is a shame, because they sound like they would have been real fun in production
- kubectl adds field manager names for help troubleshooting server-side apply conflicts
Original Source: http://lwkd.info/2020/20200519