"The Node Readiness Controller closes this gap by reconciling node readiness signals directly from the kubelet and exposing a consistent, authoritative status through the API server. In practice, this means that pods are less likely to be scheduled onto nodes experiencing transient failures, and operators gain greater confidence that scheduling decisions are based on up-to-date node state."
"In large and dynamic clusters, transient node unavailability, such as brief network interruptions between the kubelet and API server, can cause stale readiness information to persist. This stale state historically led the scheduler to think a node is healthy when it is not, resulting in pods being placed on nodes that cannot reliably start or run workloads."
"The new controller builds on Kubernetes' existing readiness mechanisms but introduces a dedicated control loop that ensures the API server's node conditions reflect the most recent and accurate health signals. By aligning API server state with actual node readiness, the feature is expected to reduce unnecessary scale-ups/spin-ups and minimize disruptive evictions triggered by outdated conditions."
Kubernetes announced the Node Readiness Controller, an alpha-stage feature addressing scheduling reliability issues in large, dynamic clusters. The controller reconciles node readiness signals directly from the kubelet and exposes consistent, authoritative status through the API server. Previously, transient network interruptions between kubelet and API server caused stale readiness information, leading the scheduler to place pods on unhealthy nodes. The new controller establishes a dedicated control loop ensuring API server node conditions reflect current health signals with reduced latency. This prevents unnecessary pod evictions, improves workload stability, and reduces false scale-ups. The feature integrates with existing mechanisms like taints, tolerations, Pod Disruption Budgets, and cluster autoscalers.
Read at InfoQ
Unable to calculate read time
Collection
[
|
...
]