Repository navigation
Graceful Node Shutdown: what it should mean, and how workload controllers find out #127
danbruno101
started this conversation in
Maintainer Summit NA 2026
Replies: 1 comment
|
This session has been scheduled for 3:05 PM: https://events.linuxfoundation.org/kubecon-cloudnativecon-north-america/features-add-ons/maintainer-summit/?id=1318482#schedule Thank you for leading this discussion! |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Topic
Graceful Node Shutdown: what it should mean, and how workload controllers find out
What projects are involved?
Graceful Node Shutdown: what it should mean, and how workload controllers find out
Description
Description
Graceful Node Shutdown: what it should mean, and how workload controllers find out
What projects are involved?
Kubernetes: SIG Node, SIG Apps, SIG Cloud Provider, SIG Windows, WG Node Lifecycle
Target Audience & Relevant Projects
Problem Statement
Graceful Node Shutdown (kubernetes/enhancements#2000) has been beta and on by default since v1.21; priority-based shutdown (kubernetes/enhancements#2712) since v1.24. Neither has moved since. Writing the follow-up KEPs this year showed that the semantics were never settled and that nothing outside the node can see a shutdown happening:
shutdownGracePeriodByPodPriorityends Pods in tiers, lowest priority first.Admit()rejects every new Pod for the whole window, on Linux and Windows. A system-critical Pod that needs to be created during the user-workload tier is rejected like everything else (Graceful Node Shutdown should allow admission of high priority pods kubernetes/kubernetes#142521).GracefulNodeShutdownInProgressyet, so controllers, autoscalers and drain tooling have no signal to read.Proposed Discussion Topic
GracefulNodeShutdownInProgressand publishing the priority boundary it is terminating at; a DaemonSet controller reader that holds back Pods below that boundary (SLM: Publish Graceful Node Shutdown state for DaemonSet coordination kubernetes/enhancements#6249); GNS as anEvictionRequestrequester (EvictionRequest API kubernetes/enhancements#4563) as the graduation step. The boundary needs a carrier, and there are two paths:PodDeletionPreference. The kubelet writes apriorityBelowrule that any controller can read, so the published API is not GNS-specific and the same object serves scale-down scoring later. Depends on the PDP KEP landing with a priority-threshold rule and node scoping.HighestPriorityAllowedfield). No new dependency, but SIG Node carries a GNS-only Node API and every consumer learns a GNS-specific field.Desired Outcomes
Session lead contact information
Session lead contact information
@danbruno101: Daniel Bruno
@rthallisey: Ryan Hallisey
@atiratree: Filip Křepinský
All reactions