kubernetes-sigs/node-feature-discovery v0.19.0

on GitHub

Highlights

• Immediate node re-labeling after a node rebuild (#2545): nfd-master detects a recreated Node via the surviving NodeFeature's owner-reference / pod-UID annotation and re-labels immediately instead of waiting for the periodic reconcile.
• Configurable NodeFeature owner references: -owner-refs / core.ownerRefs (pod, ds, node; default pod,ds) (#2545).
• core.noPublishFeatures to slim published NodeFeature objects (#2530; whitespace-trim fix #2535).
• New x86-64 microarchitecture level labels cpu-cpuid.X86_64_V1..V4 (#2490).
• nfd-topology-updater: per-pod lookups served from a node-scoped Pod informer, eliminating high-volume per-pod GETs at scale (#2509). Requires an RBAC update — see Action Required.
• nfd-master reconciles only on NodeFeature/NodeFeatureRule/NodeFeatureGroup spec changes (#2532); MatchInRegexp pattern compilation cached (#2531).

Action Required / Behavior Changes

1. nfd-topology-updater RBAC (#2509): the ClusterRole needs list+watch on pods (was get only). Helm and kustomize manifests ship the update, but deployments that bump only the image will crash-loop with pods is forbidden ... cannot list resource "pods" until the new RBAC is applied.
2. NodeFeatureRule templates (#2507): env, expandenv and getHostByName were removed from the available template functions (security hardening). Templates using them now fail to parse (function "env" not defined); the rule is skipped and an error is logged.
3. Helm worker updateStrategy default (#2519): now rollingUpdate: {maxUnavailable: "10%"} (was Kubernetes default 1). Note for type: OnDelete users: the chart's default rollingUpdate block is deep-merged into your values; the API server accepts and ignores it (verified on v1.30). Optionally set worker.updateStrategy.rollingUpdate: null for a clean spec.
4. resyncPeriod (nfd-master): drives both the informer resync and the (new in this release, #2415) periodic full reconcile that re-applies labels and cleans up drift. Setting it to 0 disables this safety net entirely.
5. Node-rebuild re-labeling dependency (#2545): immediate re-label requires pod in -owner-refs (default) or the POD_UID env set on the worker; with ds-only/empty owner refs and no POD_UID, re-labeling falls back to the periodic reconcile interval. -owner-refs=node is unaffected.

Fixes

• nfd-master: don't remove node labels when the NodeFeature cache is incomplete; add periodic reconciliation to clean up genuinely absent NodeFeatures (#2415).
• nfd-master: prevent stop() panic and informer leak on init failure (#2534).
• nfd-worker: trim whitespace in comma-separated command-line list values (#2535).
• Helm: render post-delete prune job resources at container level so they actually apply (#2515).

Other

• Container images cross-compile from $BUILDPLATFORM via tonistiigi/xx (#2512).
• Reduced cognitive complexity refactors in nfd-master controller and nfd-worker config (#2536); nfd-master command-line unit tests (#2469); codecov upload non-fatal in CI verify (#2527); template/custom-api input-validation hardening (#2507).