Unconference Session: Node Lifecycle State Needs a Real API
summary
Claim: node lifecycle state needs a structured, discoverable API because current hacks like labels and annotations and ad hoc CRDs are untyped, unowned, and break reuse. Proposals include an eviction request API (demoed and targeted for 1.37) plus node readiness gates and graceful shutdown hooks to enable in-place drain/suspend, checkpointing for GPUs, and clearer ownership. Work is prototyped and in progress; dozens of ecosystem projects already reinvent this today.
AI-written summary · about
chapters (11)
- 0:00Node lifecycle challenges
- 4:38Current node lifecycle APIs
- 7:07In-place drain and eviction
- 10:38Software dependency lifecycle
- 17:23Node readiness and transitions
- 20:11Ownership and separation
- 24:03Minimal viable primitives
- 27:16Run levels and conventions
- 33:11Events over states
- 35:58Ownership and discoverability
- 36:18Summary and collaboration
related talks
38mNode Lifecycle WorkingGroup - Getting Started
26mWriting a Custom Controller: Extending the Functionality of Your Cluster [I]
21mIntroducing The API Lifecycle
49mBuilding Software Better: API Lifecycle Management for Developers
06mSponsored Keynote: Recursive Kubernetes: Cluster API and Clusters as Cattle
25mWhat goes down the drain?
22mHave you lost your node...again?
32mInto the Deep Waters of API Machinery